1 /* 2 * Copyright (c) 2000, 2010, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "compiler/compileLog.hpp" 27 #include "memory/allocation.inline.hpp" 28 #include "opto/addnode.hpp" 29 #include "opto/callnode.hpp" 30 #include "opto/connode.hpp" 31 #include "opto/divnode.hpp" 32 #include "opto/loopnode.hpp" 33 #include "opto/mulnode.hpp" 34 #include "opto/rootnode.hpp" 35 #include "opto/runtime.hpp" 36 #include "opto/subnode.hpp" 37 38 //------------------------------is_loop_exit----------------------------------- 39 // Given an IfNode, return the loop-exiting projection or NULL if both 40 // arms remain in the loop. 41 Node *IdealLoopTree::is_loop_exit(Node *iff) const { 42 if( iff->outcnt() != 2 ) return NULL; // Ignore partially dead tests 43 PhaseIdealLoop *phase = _phase; 44 // Test is an IfNode, has 2 projections. If BOTH are in the loop 45 // we need loop unswitching instead of peeling. 46 if( !is_member(phase->get_loop( iff->raw_out(0) )) ) 47 return iff->raw_out(0); 48 if( !is_member(phase->get_loop( iff->raw_out(1) )) ) 49 return iff->raw_out(1); 50 return NULL; 51 } 52 53 54 //============================================================================= 55 56 57 //------------------------------record_for_igvn---------------------------- 58 // Put loop body on igvn work list 59 void IdealLoopTree::record_for_igvn() { 60 for( uint i = 0; i < _body.size(); i++ ) { 61 Node *n = _body.at(i); 62 _phase->_igvn._worklist.push(n); 63 } 64 } 65 66 //------------------------------compute_profile_trip_cnt---------------------------- 67 // Compute loop trip count from profile data as 68 // (backedge_count + loop_exit_count) / loop_exit_count 69 void IdealLoopTree::compute_profile_trip_cnt( PhaseIdealLoop *phase ) { 70 if (!_head->is_CountedLoop()) { 71 return; 72 } 73 CountedLoopNode* head = _head->as_CountedLoop(); 74 if (head->profile_trip_cnt() != COUNT_UNKNOWN) { 75 return; // Already computed 76 } 77 float trip_cnt = (float)max_jint; // default is big 78 79 Node* back = head->in(LoopNode::LoopBackControl); 80 while (back != head) { 81 if ((back->Opcode() == Op_IfTrue || back->Opcode() == Op_IfFalse) && 82 back->in(0) && 83 back->in(0)->is_If() && 84 back->in(0)->as_If()->_fcnt != COUNT_UNKNOWN && 85 back->in(0)->as_If()->_prob != PROB_UNKNOWN) { 86 break; 87 } 88 back = phase->idom(back); 89 } 90 if (back != head) { 91 assert((back->Opcode() == Op_IfTrue || back->Opcode() == Op_IfFalse) && 92 back->in(0), "if-projection exists"); 93 IfNode* back_if = back->in(0)->as_If(); 94 float loop_back_cnt = back_if->_fcnt * back_if->_prob; 95 96 // Now compute a loop exit count 97 float loop_exit_cnt = 0.0f; 98 for( uint i = 0; i < _body.size(); i++ ) { 99 Node *n = _body[i]; 100 if( n->is_If() ) { 101 IfNode *iff = n->as_If(); 102 if( iff->_fcnt != COUNT_UNKNOWN && iff->_prob != PROB_UNKNOWN ) { 103 Node *exit = is_loop_exit(iff); 104 if( exit ) { 105 float exit_prob = iff->_prob; 106 if (exit->Opcode() == Op_IfFalse) exit_prob = 1.0 - exit_prob; 107 if (exit_prob > PROB_MIN) { 108 float exit_cnt = iff->_fcnt * exit_prob; 109 loop_exit_cnt += exit_cnt; 110 } 111 } 112 } 113 } 114 } 115 if (loop_exit_cnt > 0.0f) { 116 trip_cnt = (loop_back_cnt + loop_exit_cnt) / loop_exit_cnt; 117 } else { 118 // No exit count so use 119 trip_cnt = loop_back_cnt; 120 } 121 } 122 #ifndef PRODUCT 123 if (TraceProfileTripCount) { 124 tty->print_cr("compute_profile_trip_cnt lp: %d cnt: %f\n", head->_idx, trip_cnt); 125 } 126 #endif 127 head->set_profile_trip_cnt(trip_cnt); 128 } 129 130 //---------------------is_invariant_addition----------------------------- 131 // Return nonzero index of invariant operand for an Add or Sub 132 // of (nonconstant) invariant and variant values. Helper for reassociate_invariants. 133 int IdealLoopTree::is_invariant_addition(Node* n, PhaseIdealLoop *phase) { 134 int op = n->Opcode(); 135 if (op == Op_AddI || op == Op_SubI) { 136 bool in1_invar = this->is_invariant(n->in(1)); 137 bool in2_invar = this->is_invariant(n->in(2)); 138 if (in1_invar && !in2_invar) return 1; 139 if (!in1_invar && in2_invar) return 2; 140 } 141 return 0; 142 } 143 144 //---------------------reassociate_add_sub----------------------------- 145 // Reassociate invariant add and subtract expressions: 146 // 147 // inv1 + (x + inv2) => ( inv1 + inv2) + x 148 // (x + inv2) + inv1 => ( inv1 + inv2) + x 149 // inv1 + (x - inv2) => ( inv1 - inv2) + x 150 // inv1 - (inv2 - x) => ( inv1 - inv2) + x 151 // (x + inv2) - inv1 => (-inv1 + inv2) + x 152 // (x - inv2) + inv1 => ( inv1 - inv2) + x 153 // (x - inv2) - inv1 => (-inv1 - inv2) + x 154 // inv1 + (inv2 - x) => ( inv1 + inv2) - x 155 // inv1 - (x - inv2) => ( inv1 + inv2) - x 156 // (inv2 - x) + inv1 => ( inv1 + inv2) - x 157 // (inv2 - x) - inv1 => (-inv1 + inv2) - x 158 // inv1 - (x + inv2) => ( inv1 - inv2) - x 159 // 160 Node* IdealLoopTree::reassociate_add_sub(Node* n1, PhaseIdealLoop *phase) { 161 if (!n1->is_Add() && !n1->is_Sub() || n1->outcnt() == 0) return NULL; 162 if (is_invariant(n1)) return NULL; 163 int inv1_idx = is_invariant_addition(n1, phase); 164 if (!inv1_idx) return NULL; 165 // Don't mess with add of constant (igvn moves them to expression tree root.) 166 if (n1->is_Add() && n1->in(2)->is_Con()) return NULL; 167 Node* inv1 = n1->in(inv1_idx); 168 Node* n2 = n1->in(3 - inv1_idx); 169 int inv2_idx = is_invariant_addition(n2, phase); 170 if (!inv2_idx) return NULL; 171 Node* x = n2->in(3 - inv2_idx); 172 Node* inv2 = n2->in(inv2_idx); 173 174 bool neg_x = n2->is_Sub() && inv2_idx == 1; 175 bool neg_inv2 = n2->is_Sub() && inv2_idx == 2; 176 bool neg_inv1 = n1->is_Sub() && inv1_idx == 2; 177 if (n1->is_Sub() && inv1_idx == 1) { 178 neg_x = !neg_x; 179 neg_inv2 = !neg_inv2; 180 } 181 Node* inv1_c = phase->get_ctrl(inv1); 182 Node* inv2_c = phase->get_ctrl(inv2); 183 Node* n_inv1; 184 if (neg_inv1) { 185 Node *zero = phase->_igvn.intcon(0); 186 phase->set_ctrl(zero, phase->C->root()); 187 n_inv1 = new (phase->C, 3) SubINode(zero, inv1); 188 phase->register_new_node(n_inv1, inv1_c); 189 } else { 190 n_inv1 = inv1; 191 } 192 Node* inv; 193 if (neg_inv2) { 194 inv = new (phase->C, 3) SubINode(n_inv1, inv2); 195 } else { 196 inv = new (phase->C, 3) AddINode(n_inv1, inv2); 197 } 198 phase->register_new_node(inv, phase->get_early_ctrl(inv)); 199 200 Node* addx; 201 if (neg_x) { 202 addx = new (phase->C, 3) SubINode(inv, x); 203 } else { 204 addx = new (phase->C, 3) AddINode(x, inv); 205 } 206 phase->register_new_node(addx, phase->get_ctrl(x)); 207 phase->_igvn.replace_node(n1, addx); 208 return addx; 209 } 210 211 //---------------------reassociate_invariants----------------------------- 212 // Reassociate invariant expressions: 213 void IdealLoopTree::reassociate_invariants(PhaseIdealLoop *phase) { 214 for (int i = _body.size() - 1; i >= 0; i--) { 215 Node *n = _body.at(i); 216 for (int j = 0; j < 5; j++) { 217 Node* nn = reassociate_add_sub(n, phase); 218 if (nn == NULL) break; 219 n = nn; // again 220 }; 221 } 222 } 223 224 //------------------------------policy_peeling--------------------------------- 225 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can 226 // make some loop-invariant test (usually a null-check) happen before the loop. 227 bool IdealLoopTree::policy_peeling( PhaseIdealLoop *phase ) const { 228 Node *test = ((IdealLoopTree*)this)->tail(); 229 int body_size = ((IdealLoopTree*)this)->_body.size(); 230 int uniq = phase->C->unique(); 231 // Peeling does loop cloning which can result in O(N^2) node construction 232 if( body_size > 255 /* Prevent overflow for large body_size */ 233 || (body_size * body_size + uniq > MaxNodeLimit) ) { 234 return false; // too large to safely clone 235 } 236 while( test != _head ) { // Scan till run off top of loop 237 if( test->is_If() ) { // Test? 238 Node *ctrl = phase->get_ctrl(test->in(1)); 239 if (ctrl->is_top()) 240 return false; // Found dead test on live IF? No peeling! 241 // Standard IF only has one input value to check for loop invariance 242 assert( test->Opcode() == Op_If || test->Opcode() == Op_CountedLoopEnd, "Check this code when new subtype is added"); 243 // Condition is not a member of this loop? 244 if( !is_member(phase->get_loop(ctrl)) && 245 is_loop_exit(test) ) 246 return true; // Found reason to peel! 247 } 248 // Walk up dominators to loop _head looking for test which is 249 // executed on every path thru loop. 250 test = phase->idom(test); 251 } 252 return false; 253 } 254 255 //------------------------------peeled_dom_test_elim--------------------------- 256 // If we got the effect of peeling, either by actually peeling or by making 257 // a pre-loop which must execute at least once, we can remove all 258 // loop-invariant dominated tests in the main body. 259 void PhaseIdealLoop::peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new ) { 260 bool progress = true; 261 while( progress ) { 262 progress = false; // Reset for next iteration 263 Node *prev = loop->_head->in(LoopNode::LoopBackControl);//loop->tail(); 264 Node *test = prev->in(0); 265 while( test != loop->_head ) { // Scan till run off top of loop 266 267 int p_op = prev->Opcode(); 268 if( (p_op == Op_IfFalse || p_op == Op_IfTrue) && 269 test->is_If() && // Test? 270 !test->in(1)->is_Con() && // And not already obvious? 271 // Condition is not a member of this loop? 272 !loop->is_member(get_loop(get_ctrl(test->in(1))))){ 273 // Walk loop body looking for instances of this test 274 for( uint i = 0; i < loop->_body.size(); i++ ) { 275 Node *n = loop->_body.at(i); 276 if( n->is_If() && n->in(1) == test->in(1) /*&& n != loop->tail()->in(0)*/ ) { 277 // IfNode was dominated by version in peeled loop body 278 progress = true; 279 dominated_by( old_new[prev->_idx], n ); 280 } 281 } 282 } 283 prev = test; 284 test = idom(test); 285 } // End of scan tests in loop 286 287 } // End of while( progress ) 288 } 289 290 //------------------------------do_peeling------------------------------------- 291 // Peel the first iteration of the given loop. 292 // Step 1: Clone the loop body. The clone becomes the peeled iteration. 293 // The pre-loop illegally has 2 control users (old & new loops). 294 // Step 2: Make the old-loop fall-in edges point to the peeled iteration. 295 // Do this by making the old-loop fall-in edges act as if they came 296 // around the loopback from the prior iteration (follow the old-loop 297 // backedges) and then map to the new peeled iteration. This leaves 298 // the pre-loop with only 1 user (the new peeled iteration), but the 299 // peeled-loop backedge has 2 users. 300 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the 301 // extra backedge user. 302 void PhaseIdealLoop::do_peeling( IdealLoopTree *loop, Node_List &old_new ) { 303 304 C->set_major_progress(); 305 // Peeling a 'main' loop in a pre/main/post situation obfuscates the 306 // 'pre' loop from the main and the 'pre' can no longer have it's 307 // iterations adjusted. Therefore, we need to declare this loop as 308 // no longer a 'main' loop; it will need new pre and post loops before 309 // we can do further RCE. 310 Node *h = loop->_head; 311 if( h->is_CountedLoop() ) { 312 CountedLoopNode *cl = h->as_CountedLoop(); 313 assert(cl->trip_count() > 0, "peeling a fully unrolled loop"); 314 cl->set_trip_count(cl->trip_count() - 1); 315 if( cl->is_main_loop() ) { 316 cl->set_normal_loop(); 317 #ifndef PRODUCT 318 if( PrintOpto && VerifyLoopOptimizations ) { 319 tty->print("Peeling a 'main' loop; resetting to 'normal' "); 320 loop->dump_head(); 321 } 322 #endif 323 } 324 } 325 326 // Step 1: Clone the loop body. The clone becomes the peeled iteration. 327 // The pre-loop illegally has 2 control users (old & new loops). 328 clone_loop( loop, old_new, dom_depth(loop->_head) ); 329 330 331 // Step 2: Make the old-loop fall-in edges point to the peeled iteration. 332 // Do this by making the old-loop fall-in edges act as if they came 333 // around the loopback from the prior iteration (follow the old-loop 334 // backedges) and then map to the new peeled iteration. This leaves 335 // the pre-loop with only 1 user (the new peeled iteration), but the 336 // peeled-loop backedge has 2 users. 337 for (DUIterator_Fast jmax, j = loop->_head->fast_outs(jmax); j < jmax; j++) { 338 Node* old = loop->_head->fast_out(j); 339 if( old->in(0) == loop->_head && old->req() == 3 && 340 (old->is_Loop() || old->is_Phi()) ) { 341 Node *new_exit_value = old_new[old->in(LoopNode::LoopBackControl)->_idx]; 342 if( !new_exit_value ) // Backedge value is ALSO loop invariant? 343 // Then loop body backedge value remains the same. 344 new_exit_value = old->in(LoopNode::LoopBackControl); 345 _igvn.hash_delete(old); 346 old->set_req(LoopNode::EntryControl, new_exit_value); 347 } 348 } 349 350 351 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the 352 // extra backedge user. 353 Node *nnn = old_new[loop->_head->_idx]; 354 _igvn.hash_delete(nnn); 355 nnn->set_req(LoopNode::LoopBackControl, C->top()); 356 for (DUIterator_Fast j2max, j2 = nnn->fast_outs(j2max); j2 < j2max; j2++) { 357 Node* use = nnn->fast_out(j2); 358 if( use->in(0) == nnn && use->req() == 3 && use->is_Phi() ) { 359 _igvn.hash_delete(use); 360 use->set_req(LoopNode::LoopBackControl, C->top()); 361 } 362 } 363 364 365 // Step 4: Correct dom-depth info. Set to loop-head depth. 366 int dd = dom_depth(loop->_head); 367 set_idom(loop->_head, loop->_head->in(1), dd); 368 for (uint j3 = 0; j3 < loop->_body.size(); j3++) { 369 Node *old = loop->_body.at(j3); 370 Node *nnn = old_new[old->_idx]; 371 if (!has_ctrl(nnn)) 372 set_idom(nnn, idom(nnn), dd-1); 373 // While we're at it, remove any SafePoints from the peeled code 374 if( old->Opcode() == Op_SafePoint ) { 375 Node *nnn = old_new[old->_idx]; 376 lazy_replace(nnn,nnn->in(TypeFunc::Control)); 377 } 378 } 379 380 // Now force out all loop-invariant dominating tests. The optimizer 381 // finds some, but we _know_ they are all useless. 382 peeled_dom_test_elim(loop,old_new); 383 384 loop->record_for_igvn(); 385 } 386 387 //------------------------------policy_maximally_unroll------------------------ 388 // Return exact loop trip count, or 0 if not maximally unrolling 389 bool IdealLoopTree::policy_maximally_unroll( PhaseIdealLoop *phase ) const { 390 CountedLoopNode *cl = _head->as_CountedLoop(); 391 assert( cl->is_normal_loop(), "" ); 392 393 Node *init_n = cl->init_trip(); 394 Node *limit_n = cl->limit(); 395 396 // Non-constant bounds 397 if( init_n == NULL || !init_n->is_Con() || 398 limit_n == NULL || !limit_n->is_Con() || 399 // protect against stride not being a constant 400 !cl->stride_is_con() ) { 401 return false; 402 } 403 int init = init_n->get_int(); 404 int limit = limit_n->get_int(); 405 int span = limit - init; 406 int stride = cl->stride_con(); 407 408 if (init >= limit || stride > span) { 409 // return a false (no maximally unroll) and the regular unroll/peel 410 // route will make a small mess which CCP will fold away. 411 return false; 412 } 413 uint trip_count = span/stride; // trip_count can be greater than 2 Gig. 414 assert( (int)trip_count*stride == span, "must divide evenly" ); 415 416 // Real policy: if we maximally unroll, does it get too big? 417 // Allow the unrolled mess to get larger than standard loop 418 // size. After all, it will no longer be a loop. 419 uint body_size = _body.size(); 420 uint unroll_limit = (uint)LoopUnrollLimit * 4; 421 assert( (intx)unroll_limit == LoopUnrollLimit * 4, "LoopUnrollLimit must fit in 32bits"); 422 cl->set_trip_count(trip_count); 423 if( trip_count <= unroll_limit && body_size <= unroll_limit ) { 424 uint new_body_size = body_size * trip_count; 425 if (new_body_size <= unroll_limit && 426 body_size == new_body_size / trip_count && 427 // Unrolling can result in a large amount of node construction 428 new_body_size < MaxNodeLimit - phase->C->unique()) { 429 return true; // maximally unroll 430 } 431 } 432 433 return false; // Do not maximally unroll 434 } 435 436 437 //------------------------------policy_unroll---------------------------------- 438 // Return TRUE or FALSE if the loop should be unrolled or not. Unroll if 439 // the loop is a CountedLoop and the body is small enough. 440 bool IdealLoopTree::policy_unroll( PhaseIdealLoop *phase ) const { 441 442 CountedLoopNode *cl = _head->as_CountedLoop(); 443 assert( cl->is_normal_loop() || cl->is_main_loop(), "" ); 444 445 // protect against stride not being a constant 446 if( !cl->stride_is_con() ) return false; 447 448 // protect against over-unrolling 449 if( cl->trip_count() <= 1 ) return false; 450 451 int future_unroll_ct = cl->unrolled_count() * 2; 452 453 // Don't unroll if the next round of unrolling would push us 454 // over the expected trip count of the loop. One is subtracted 455 // from the expected trip count because the pre-loop normally 456 // executes 1 iteration. 457 if (UnrollLimitForProfileCheck > 0 && 458 cl->profile_trip_cnt() != COUNT_UNKNOWN && 459 future_unroll_ct > UnrollLimitForProfileCheck && 460 (float)future_unroll_ct > cl->profile_trip_cnt() - 1.0) { 461 return false; 462 } 463 464 // When unroll count is greater than LoopUnrollMin, don't unroll if: 465 // the residual iterations are more than 10% of the trip count 466 // and rounds of "unroll,optimize" are not making significant progress 467 // Progress defined as current size less than 20% larger than previous size. 468 if (UseSuperWord && cl->node_count_before_unroll() > 0 && 469 future_unroll_ct > LoopUnrollMin && 470 (future_unroll_ct - 1) * 10.0 > cl->profile_trip_cnt() && 471 1.2 * cl->node_count_before_unroll() < (double)_body.size()) { 472 return false; 473 } 474 475 Node *init_n = cl->init_trip(); 476 Node *limit_n = cl->limit(); 477 // Non-constant bounds. 478 // Protect against over-unrolling when init or/and limit are not constant 479 // (so that trip_count's init value is maxint) but iv range is known. 480 if( init_n == NULL || !init_n->is_Con() || 481 limit_n == NULL || !limit_n->is_Con() ) { 482 Node* phi = cl->phi(); 483 if( phi != NULL ) { 484 assert(phi->is_Phi() && phi->in(0) == _head, "Counted loop should have iv phi."); 485 const TypeInt* iv_type = phase->_igvn.type(phi)->is_int(); 486 int next_stride = cl->stride_con() * 2; // stride after this unroll 487 if( next_stride > 0 ) { 488 if( iv_type->_lo + next_stride <= iv_type->_lo || // overflow 489 iv_type->_lo + next_stride > iv_type->_hi ) { 490 return false; // over-unrolling 491 } 492 } else if( next_stride < 0 ) { 493 if( iv_type->_hi + next_stride >= iv_type->_hi || // overflow 494 iv_type->_hi + next_stride < iv_type->_lo ) { 495 return false; // over-unrolling 496 } 497 } 498 } 499 } 500 501 // Adjust body_size to determine if we unroll or not 502 uint body_size = _body.size(); 503 // Key test to unroll CaffeineMark's Logic test 504 int xors_in_loop = 0; 505 // Also count ModL, DivL and MulL which expand mightly 506 for( uint k = 0; k < _body.size(); k++ ) { 507 switch( _body.at(k)->Opcode() ) { 508 case Op_XorI: xors_in_loop++; break; // CaffeineMark's Logic test 509 case Op_ModL: body_size += 30; break; 510 case Op_DivL: body_size += 30; break; 511 case Op_MulL: body_size += 10; break; 512 } 513 } 514 515 // Check for being too big 516 if( body_size > (uint)LoopUnrollLimit ) { 517 if( xors_in_loop >= 4 && body_size < (uint)LoopUnrollLimit*4) return true; 518 // Normal case: loop too big 519 return false; 520 } 521 522 // Check for stride being a small enough constant 523 if( abs(cl->stride_con()) > (1<<3) ) return false; 524 525 // Unroll once! (Each trip will soon do double iterations) 526 return true; 527 } 528 529 //------------------------------policy_align----------------------------------- 530 // Return TRUE or FALSE if the loop should be cache-line aligned. Gather the 531 // expression that does the alignment. Note that only one array base can be 532 // aligned in a loop (unless the VM guarantees mutual alignment). Note that 533 // if we vectorize short memory ops into longer memory ops, we may want to 534 // increase alignment. 535 bool IdealLoopTree::policy_align( PhaseIdealLoop *phase ) const { 536 return false; 537 } 538 539 //------------------------------policy_range_check----------------------------- 540 // Return TRUE or FALSE if the loop should be range-check-eliminated. 541 // Actually we do iteration-splitting, a more powerful form of RCE. 542 bool IdealLoopTree::policy_range_check( PhaseIdealLoop *phase ) const { 543 if( !RangeCheckElimination ) return false; 544 545 CountedLoopNode *cl = _head->as_CountedLoop(); 546 // If we unrolled with no intention of doing RCE and we later 547 // changed our minds, we got no pre-loop. Either we need to 548 // make a new pre-loop, or we gotta disallow RCE. 549 if( cl->is_main_no_pre_loop() ) return false; // Disallowed for now. 550 Node *trip_counter = cl->phi(); 551 552 // Check loop body for tests of trip-counter plus loop-invariant vs 553 // loop-invariant. 554 for( uint i = 0; i < _body.size(); i++ ) { 555 Node *iff = _body[i]; 556 if( iff->Opcode() == Op_If ) { // Test? 557 558 // Comparing trip+off vs limit 559 Node *bol = iff->in(1); 560 if( bol->req() != 2 ) continue; // dead constant test 561 if (!bol->is_Bool()) { 562 assert(UseLoopPredicate && bol->Opcode() == Op_Conv2B, "predicate check only"); 563 continue; 564 } 565 Node *cmp = bol->in(1); 566 567 Node *rc_exp = cmp->in(1); 568 Node *limit = cmp->in(2); 569 570 Node *limit_c = phase->get_ctrl(limit); 571 if( limit_c == phase->C->top() ) 572 return false; // Found dead test on live IF? No RCE! 573 if( is_member(phase->get_loop(limit_c) ) ) { 574 // Compare might have operands swapped; commute them 575 rc_exp = cmp->in(2); 576 limit = cmp->in(1); 577 limit_c = phase->get_ctrl(limit); 578 if( is_member(phase->get_loop(limit_c) ) ) 579 continue; // Both inputs are loop varying; cannot RCE 580 } 581 582 if (!phase->is_scaled_iv_plus_offset(rc_exp, trip_counter, NULL, NULL)) { 583 continue; 584 } 585 // Yeah! Found a test like 'trip+off vs limit' 586 // Test is an IfNode, has 2 projections. If BOTH are in the loop 587 // we need loop unswitching instead of iteration splitting. 588 if( is_loop_exit(iff) ) 589 return true; // Found reason to split iterations 590 } // End of is IF 591 } 592 593 return false; 594 } 595 596 //------------------------------policy_peel_only------------------------------- 597 // Return TRUE or FALSE if the loop should NEVER be RCE'd or aligned. Useful 598 // for unrolling loops with NO array accesses. 599 bool IdealLoopTree::policy_peel_only( PhaseIdealLoop *phase ) const { 600 601 for( uint i = 0; i < _body.size(); i++ ) 602 if( _body[i]->is_Mem() ) 603 return false; 604 605 // No memory accesses at all! 606 return true; 607 } 608 609 //------------------------------clone_up_backedge_goo-------------------------- 610 // If Node n lives in the back_ctrl block and cannot float, we clone a private 611 // version of n in preheader_ctrl block and return that, otherwise return n. 612 Node *PhaseIdealLoop::clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n ) { 613 if( get_ctrl(n) != back_ctrl ) return n; 614 615 Node *x = NULL; // If required, a clone of 'n' 616 // Check for 'n' being pinned in the backedge. 617 if( n->in(0) && n->in(0) == back_ctrl ) { 618 x = n->clone(); // Clone a copy of 'n' to preheader 619 x->set_req( 0, preheader_ctrl ); // Fix x's control input to preheader 620 } 621 622 // Recursive fixup any other input edges into x. 623 // If there are no changes we can just return 'n', otherwise 624 // we need to clone a private copy and change it. 625 for( uint i = 1; i < n->req(); i++ ) { 626 Node *g = clone_up_backedge_goo( back_ctrl, preheader_ctrl, n->in(i) ); 627 if( g != n->in(i) ) { 628 if( !x ) 629 x = n->clone(); 630 x->set_req(i, g); 631 } 632 } 633 if( x ) { // x can legally float to pre-header location 634 register_new_node( x, preheader_ctrl ); 635 return x; 636 } else { // raise n to cover LCA of uses 637 set_ctrl( n, find_non_split_ctrl(back_ctrl->in(0)) ); 638 } 639 return n; 640 } 641 642 //------------------------------insert_pre_post_loops-------------------------- 643 // Insert pre and post loops. If peel_only is set, the pre-loop can not have 644 // more iterations added. It acts as a 'peel' only, no lower-bound RCE, no 645 // alignment. Useful to unroll loops that do no array accesses. 646 void PhaseIdealLoop::insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only ) { 647 648 C->set_major_progress(); 649 650 // Find common pieces of the loop being guarded with pre & post loops 651 CountedLoopNode *main_head = loop->_head->as_CountedLoop(); 652 assert( main_head->is_normal_loop(), "" ); 653 CountedLoopEndNode *main_end = main_head->loopexit(); 654 assert( main_end->outcnt() == 2, "1 true, 1 false path only" ); 655 uint dd_main_head = dom_depth(main_head); 656 uint max = main_head->outcnt(); 657 658 Node *pre_header= main_head->in(LoopNode::EntryControl); 659 Node *init = main_head->init_trip(); 660 Node *incr = main_end ->incr(); 661 Node *limit = main_end ->limit(); 662 Node *stride = main_end ->stride(); 663 Node *cmp = main_end ->cmp_node(); 664 BoolTest::mask b_test = main_end->test_trip(); 665 666 // Need only 1 user of 'bol' because I will be hacking the loop bounds. 667 Node *bol = main_end->in(CountedLoopEndNode::TestValue); 668 if( bol->outcnt() != 1 ) { 669 bol = bol->clone(); 670 register_new_node(bol,main_end->in(CountedLoopEndNode::TestControl)); 671 _igvn.hash_delete(main_end); 672 main_end->set_req(CountedLoopEndNode::TestValue, bol); 673 } 674 // Need only 1 user of 'cmp' because I will be hacking the loop bounds. 675 if( cmp->outcnt() != 1 ) { 676 cmp = cmp->clone(); 677 register_new_node(cmp,main_end->in(CountedLoopEndNode::TestControl)); 678 _igvn.hash_delete(bol); 679 bol->set_req(1, cmp); 680 } 681 682 //------------------------------ 683 // Step A: Create Post-Loop. 684 Node* main_exit = main_end->proj_out(false); 685 assert( main_exit->Opcode() == Op_IfFalse, "" ); 686 int dd_main_exit = dom_depth(main_exit); 687 688 // Step A1: Clone the loop body. The clone becomes the post-loop. The main 689 // loop pre-header illegally has 2 control users (old & new loops). 690 clone_loop( loop, old_new, dd_main_exit ); 691 assert( old_new[main_end ->_idx]->Opcode() == Op_CountedLoopEnd, "" ); 692 CountedLoopNode *post_head = old_new[main_head->_idx]->as_CountedLoop(); 693 post_head->set_post_loop(main_head); 694 695 // Reduce the post-loop trip count. 696 CountedLoopEndNode* post_end = old_new[main_end ->_idx]->as_CountedLoopEnd(); 697 post_end->_prob = PROB_FAIR; 698 699 // Build the main-loop normal exit. 700 IfFalseNode *new_main_exit = new (C, 1) IfFalseNode(main_end); 701 _igvn.register_new_node_with_optimizer( new_main_exit ); 702 set_idom(new_main_exit, main_end, dd_main_exit ); 703 set_loop(new_main_exit, loop->_parent); 704 705 // Step A2: Build a zero-trip guard for the post-loop. After leaving the 706 // main-loop, the post-loop may not execute at all. We 'opaque' the incr 707 // (the main-loop trip-counter exit value) because we will be changing 708 // the exit value (via unrolling) so we cannot constant-fold away the zero 709 // trip guard until all unrolling is done. 710 Node *zer_opaq = new (C, 2) Opaque1Node(C, incr); 711 Node *zer_cmp = new (C, 3) CmpINode( zer_opaq, limit ); 712 Node *zer_bol = new (C, 2) BoolNode( zer_cmp, b_test ); 713 register_new_node( zer_opaq, new_main_exit ); 714 register_new_node( zer_cmp , new_main_exit ); 715 register_new_node( zer_bol , new_main_exit ); 716 717 // Build the IfNode 718 IfNode *zer_iff = new (C, 2) IfNode( new_main_exit, zer_bol, PROB_FAIR, COUNT_UNKNOWN ); 719 _igvn.register_new_node_with_optimizer( zer_iff ); 720 set_idom(zer_iff, new_main_exit, dd_main_exit); 721 set_loop(zer_iff, loop->_parent); 722 723 // Plug in the false-path, taken if we need to skip post-loop 724 _igvn.hash_delete( main_exit ); 725 main_exit->set_req(0, zer_iff); 726 _igvn._worklist.push(main_exit); 727 set_idom(main_exit, zer_iff, dd_main_exit); 728 set_idom(main_exit->unique_out(), zer_iff, dd_main_exit); 729 // Make the true-path, must enter the post loop 730 Node *zer_taken = new (C, 1) IfTrueNode( zer_iff ); 731 _igvn.register_new_node_with_optimizer( zer_taken ); 732 set_idom(zer_taken, zer_iff, dd_main_exit); 733 set_loop(zer_taken, loop->_parent); 734 // Plug in the true path 735 _igvn.hash_delete( post_head ); 736 post_head->set_req(LoopNode::EntryControl, zer_taken); 737 set_idom(post_head, zer_taken, dd_main_exit); 738 739 // Step A3: Make the fall-in values to the post-loop come from the 740 // fall-out values of the main-loop. 741 for (DUIterator_Fast imax, i = main_head->fast_outs(imax); i < imax; i++) { 742 Node* main_phi = main_head->fast_out(i); 743 if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() >0 ) { 744 Node *post_phi = old_new[main_phi->_idx]; 745 Node *fallmain = clone_up_backedge_goo(main_head->back_control(), 746 post_head->init_control(), 747 main_phi->in(LoopNode::LoopBackControl)); 748 _igvn.hash_delete(post_phi); 749 post_phi->set_req( LoopNode::EntryControl, fallmain ); 750 } 751 } 752 753 // Update local caches for next stanza 754 main_exit = new_main_exit; 755 756 757 //------------------------------ 758 // Step B: Create Pre-Loop. 759 760 // Step B1: Clone the loop body. The clone becomes the pre-loop. The main 761 // loop pre-header illegally has 2 control users (old & new loops). 762 clone_loop( loop, old_new, dd_main_head ); 763 CountedLoopNode* pre_head = old_new[main_head->_idx]->as_CountedLoop(); 764 CountedLoopEndNode* pre_end = old_new[main_end ->_idx]->as_CountedLoopEnd(); 765 pre_head->set_pre_loop(main_head); 766 Node *pre_incr = old_new[incr->_idx]; 767 768 // Reduce the pre-loop trip count. 769 pre_end->_prob = PROB_FAIR; 770 771 // Find the pre-loop normal exit. 772 Node* pre_exit = pre_end->proj_out(false); 773 assert( pre_exit->Opcode() == Op_IfFalse, "" ); 774 IfFalseNode *new_pre_exit = new (C, 1) IfFalseNode(pre_end); 775 _igvn.register_new_node_with_optimizer( new_pre_exit ); 776 set_idom(new_pre_exit, pre_end, dd_main_head); 777 set_loop(new_pre_exit, loop->_parent); 778 779 // Step B2: Build a zero-trip guard for the main-loop. After leaving the 780 // pre-loop, the main-loop may not execute at all. Later in life this 781 // zero-trip guard will become the minimum-trip guard when we unroll 782 // the main-loop. 783 Node *min_opaq = new (C, 2) Opaque1Node(C, limit); 784 Node *min_cmp = new (C, 3) CmpINode( pre_incr, min_opaq ); 785 Node *min_bol = new (C, 2) BoolNode( min_cmp, b_test ); 786 register_new_node( min_opaq, new_pre_exit ); 787 register_new_node( min_cmp , new_pre_exit ); 788 register_new_node( min_bol , new_pre_exit ); 789 790 // Build the IfNode (assume the main-loop is executed always). 791 IfNode *min_iff = new (C, 2) IfNode( new_pre_exit, min_bol, PROB_ALWAYS, COUNT_UNKNOWN ); 792 _igvn.register_new_node_with_optimizer( min_iff ); 793 set_idom(min_iff, new_pre_exit, dd_main_head); 794 set_loop(min_iff, loop->_parent); 795 796 // Plug in the false-path, taken if we need to skip main-loop 797 _igvn.hash_delete( pre_exit ); 798 pre_exit->set_req(0, min_iff); 799 set_idom(pre_exit, min_iff, dd_main_head); 800 set_idom(pre_exit->unique_out(), min_iff, dd_main_head); 801 // Make the true-path, must enter the main loop 802 Node *min_taken = new (C, 1) IfTrueNode( min_iff ); 803 _igvn.register_new_node_with_optimizer( min_taken ); 804 set_idom(min_taken, min_iff, dd_main_head); 805 set_loop(min_taken, loop->_parent); 806 // Plug in the true path 807 _igvn.hash_delete( main_head ); 808 main_head->set_req(LoopNode::EntryControl, min_taken); 809 set_idom(main_head, min_taken, dd_main_head); 810 811 // Step B3: Make the fall-in values to the main-loop come from the 812 // fall-out values of the pre-loop. 813 for (DUIterator_Fast i2max, i2 = main_head->fast_outs(i2max); i2 < i2max; i2++) { 814 Node* main_phi = main_head->fast_out(i2); 815 if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() > 0 ) { 816 Node *pre_phi = old_new[main_phi->_idx]; 817 Node *fallpre = clone_up_backedge_goo(pre_head->back_control(), 818 main_head->init_control(), 819 pre_phi->in(LoopNode::LoopBackControl)); 820 _igvn.hash_delete(main_phi); 821 main_phi->set_req( LoopNode::EntryControl, fallpre ); 822 } 823 } 824 825 // Step B4: Shorten the pre-loop to run only 1 iteration (for now). 826 // RCE and alignment may change this later. 827 Node *cmp_end = pre_end->cmp_node(); 828 assert( cmp_end->in(2) == limit, "" ); 829 Node *pre_limit = new (C, 3) AddINode( init, stride ); 830 831 // Save the original loop limit in this Opaque1 node for 832 // use by range check elimination. 833 Node *pre_opaq = new (C, 3) Opaque1Node(C, pre_limit, limit); 834 835 register_new_node( pre_limit, pre_head->in(0) ); 836 register_new_node( pre_opaq , pre_head->in(0) ); 837 838 // Since no other users of pre-loop compare, I can hack limit directly 839 assert( cmp_end->outcnt() == 1, "no other users" ); 840 _igvn.hash_delete(cmp_end); 841 cmp_end->set_req(2, peel_only ? pre_limit : pre_opaq); 842 843 // Special case for not-equal loop bounds: 844 // Change pre loop test, main loop test, and the 845 // main loop guard test to use lt or gt depending on stride 846 // direction: 847 // positive stride use < 848 // negative stride use > 849 850 if (pre_end->in(CountedLoopEndNode::TestValue)->as_Bool()->_test._test == BoolTest::ne) { 851 852 BoolTest::mask new_test = (main_end->stride_con() > 0) ? BoolTest::lt : BoolTest::gt; 853 // Modify pre loop end condition 854 Node* pre_bol = pre_end->in(CountedLoopEndNode::TestValue)->as_Bool(); 855 BoolNode* new_bol0 = new (C, 2) BoolNode(pre_bol->in(1), new_test); 856 register_new_node( new_bol0, pre_head->in(0) ); 857 _igvn.hash_delete(pre_end); 858 pre_end->set_req(CountedLoopEndNode::TestValue, new_bol0); 859 // Modify main loop guard condition 860 assert(min_iff->in(CountedLoopEndNode::TestValue) == min_bol, "guard okay"); 861 BoolNode* new_bol1 = new (C, 2) BoolNode(min_bol->in(1), new_test); 862 register_new_node( new_bol1, new_pre_exit ); 863 _igvn.hash_delete(min_iff); 864 min_iff->set_req(CountedLoopEndNode::TestValue, new_bol1); 865 // Modify main loop end condition 866 BoolNode* main_bol = main_end->in(CountedLoopEndNode::TestValue)->as_Bool(); 867 BoolNode* new_bol2 = new (C, 2) BoolNode(main_bol->in(1), new_test); 868 register_new_node( new_bol2, main_end->in(CountedLoopEndNode::TestControl) ); 869 _igvn.hash_delete(main_end); 870 main_end->set_req(CountedLoopEndNode::TestValue, new_bol2); 871 } 872 873 // Flag main loop 874 main_head->set_main_loop(); 875 if( peel_only ) main_head->set_main_no_pre_loop(); 876 877 // It's difficult to be precise about the trip-counts 878 // for the pre/post loops. They are usually very short, 879 // so guess that 4 trips is a reasonable value. 880 post_head->set_profile_trip_cnt(4.0); 881 pre_head->set_profile_trip_cnt(4.0); 882 883 // Now force out all loop-invariant dominating tests. The optimizer 884 // finds some, but we _know_ they are all useless. 885 peeled_dom_test_elim(loop,old_new); 886 } 887 888 //------------------------------is_invariant----------------------------- 889 // Return true if n is invariant 890 bool IdealLoopTree::is_invariant(Node* n) const { 891 Node *n_c = _phase->has_ctrl(n) ? _phase->get_ctrl(n) : n; 892 if (n_c->is_top()) return false; 893 return !is_member(_phase->get_loop(n_c)); 894 } 895 896 897 //------------------------------do_unroll-------------------------------------- 898 // Unroll the loop body one step - make each trip do 2 iterations. 899 void PhaseIdealLoop::do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip ) { 900 assert( LoopUnrollLimit, "" ); 901 #ifndef PRODUCT 902 if( PrintOpto && VerifyLoopOptimizations ) { 903 tty->print("Unrolling "); 904 loop->dump_head(); 905 } 906 #endif 907 CountedLoopNode *loop_head = loop->_head->as_CountedLoop(); 908 CountedLoopEndNode *loop_end = loop_head->loopexit(); 909 assert( loop_end, "" ); 910 911 // Remember loop node count before unrolling to detect 912 // if rounds of unroll,optimize are making progress 913 loop_head->set_node_count_before_unroll(loop->_body.size()); 914 915 Node *ctrl = loop_head->in(LoopNode::EntryControl); 916 Node *limit = loop_head->limit(); 917 Node *init = loop_head->init_trip(); 918 Node *strid = loop_head->stride(); 919 920 Node *opaq = NULL; 921 if( adjust_min_trip ) { // If not maximally unrolling, need adjustment 922 assert( loop_head->is_main_loop(), "" ); 923 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" ); 924 Node *iff = ctrl->in(0); 925 assert( iff->Opcode() == Op_If, "" ); 926 Node *bol = iff->in(1); 927 assert( bol->Opcode() == Op_Bool, "" ); 928 Node *cmp = bol->in(1); 929 assert( cmp->Opcode() == Op_CmpI, "" ); 930 opaq = cmp->in(2); 931 // Occasionally it's possible for a pre-loop Opaque1 node to be 932 // optimized away and then another round of loop opts attempted. 933 // We can not optimize this particular loop in that case. 934 if( opaq->Opcode() != Op_Opaque1 ) 935 return; // Cannot find pre-loop! Bail out! 936 } 937 938 C->set_major_progress(); 939 940 // Adjust max trip count. The trip count is intentionally rounded 941 // down here (e.g. 15-> 7-> 3-> 1) because if we unwittingly over-unroll, 942 // the main, unrolled, part of the loop will never execute as it is protected 943 // by the min-trip test. See bug 4834191 for a case where we over-unrolled 944 // and later determined that part of the unrolled loop was dead. 945 loop_head->set_trip_count(loop_head->trip_count() / 2); 946 947 // Double the count of original iterations in the unrolled loop body. 948 loop_head->double_unrolled_count(); 949 950 // ----------- 951 // Step 2: Cut back the trip counter for an unroll amount of 2. 952 // Loop will normally trip (limit - init)/stride_con. Since it's a 953 // CountedLoop this is exact (stride divides limit-init exactly). 954 // We are going to double the loop body, so we want to knock off any 955 // odd iteration: (trip_cnt & ~1). Then back compute a new limit. 956 Node *span = new (C, 3) SubINode( limit, init ); 957 register_new_node( span, ctrl ); 958 Node *trip = new (C, 3) DivINode( 0, span, strid ); 959 register_new_node( trip, ctrl ); 960 Node *mtwo = _igvn.intcon(-2); 961 set_ctrl(mtwo, C->root()); 962 Node *rond = new (C, 3) AndINode( trip, mtwo ); 963 register_new_node( rond, ctrl ); 964 Node *spn2 = new (C, 3) MulINode( rond, strid ); 965 register_new_node( spn2, ctrl ); 966 Node *lim2 = new (C, 3) AddINode( spn2, init ); 967 register_new_node( lim2, ctrl ); 968 969 // Hammer in the new limit 970 Node *ctrl2 = loop_end->in(0); 971 Node *cmp2 = new (C, 3) CmpINode( loop_head->incr(), lim2 ); 972 register_new_node( cmp2, ctrl2 ); 973 Node *bol2 = new (C, 2) BoolNode( cmp2, loop_end->test_trip() ); 974 register_new_node( bol2, ctrl2 ); 975 _igvn.hash_delete(loop_end); 976 loop_end->set_req(CountedLoopEndNode::TestValue, bol2); 977 978 // Step 3: Find the min-trip test guaranteed before a 'main' loop. 979 // Make it a 1-trip test (means at least 2 trips). 980 if( adjust_min_trip ) { 981 // Guard test uses an 'opaque' node which is not shared. Hence I 982 // can edit it's inputs directly. Hammer in the new limit for the 983 // minimum-trip guard. 984 assert( opaq->outcnt() == 1, "" ); 985 _igvn.hash_delete(opaq); 986 opaq->set_req(1, lim2); 987 } 988 989 // --------- 990 // Step 4: Clone the loop body. Move it inside the loop. This loop body 991 // represents the odd iterations; since the loop trips an even number of 992 // times its backedge is never taken. Kill the backedge. 993 uint dd = dom_depth(loop_head); 994 clone_loop( loop, old_new, dd ); 995 996 // Make backedges of the clone equal to backedges of the original. 997 // Make the fall-in from the original come from the fall-out of the clone. 998 for (DUIterator_Fast jmax, j = loop_head->fast_outs(jmax); j < jmax; j++) { 999 Node* phi = loop_head->fast_out(j); 1000 if( phi->is_Phi() && phi->in(0) == loop_head && phi->outcnt() > 0 ) { 1001 Node *newphi = old_new[phi->_idx]; 1002 _igvn.hash_delete( phi ); 1003 _igvn.hash_delete( newphi ); 1004 1005 phi ->set_req(LoopNode:: EntryControl, newphi->in(LoopNode::LoopBackControl)); 1006 newphi->set_req(LoopNode::LoopBackControl, phi ->in(LoopNode::LoopBackControl)); 1007 phi ->set_req(LoopNode::LoopBackControl, C->top()); 1008 } 1009 } 1010 Node *clone_head = old_new[loop_head->_idx]; 1011 _igvn.hash_delete( clone_head ); 1012 loop_head ->set_req(LoopNode:: EntryControl, clone_head->in(LoopNode::LoopBackControl)); 1013 clone_head->set_req(LoopNode::LoopBackControl, loop_head ->in(LoopNode::LoopBackControl)); 1014 loop_head ->set_req(LoopNode::LoopBackControl, C->top()); 1015 loop->_head = clone_head; // New loop header 1016 1017 set_idom(loop_head, loop_head ->in(LoopNode::EntryControl), dd); 1018 set_idom(clone_head, clone_head->in(LoopNode::EntryControl), dd); 1019 1020 // Kill the clone's backedge 1021 Node *newcle = old_new[loop_end->_idx]; 1022 _igvn.hash_delete( newcle ); 1023 Node *one = _igvn.intcon(1); 1024 set_ctrl(one, C->root()); 1025 newcle->set_req(1, one); 1026 // Force clone into same loop body 1027 uint max = loop->_body.size(); 1028 for( uint k = 0; k < max; k++ ) { 1029 Node *old = loop->_body.at(k); 1030 Node *nnn = old_new[old->_idx]; 1031 loop->_body.push(nnn); 1032 if (!has_ctrl(old)) 1033 set_loop(nnn, loop); 1034 } 1035 1036 loop->record_for_igvn(); 1037 } 1038 1039 //------------------------------do_maximally_unroll---------------------------- 1040 1041 void PhaseIdealLoop::do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new ) { 1042 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 1043 assert( cl->trip_count() > 0, ""); 1044 1045 // If loop is tripping an odd number of times, peel odd iteration 1046 if( (cl->trip_count() & 1) == 1 ) { 1047 do_peeling( loop, old_new ); 1048 } 1049 1050 // Now its tripping an even number of times remaining. Double loop body. 1051 // Do not adjust pre-guards; they are not needed and do not exist. 1052 if( cl->trip_count() > 0 ) { 1053 do_unroll( loop, old_new, false ); 1054 } 1055 } 1056 1057 //------------------------------dominates_backedge--------------------------------- 1058 // Returns true if ctrl is executed on every complete iteration 1059 bool IdealLoopTree::dominates_backedge(Node* ctrl) { 1060 assert(ctrl->is_CFG(), "must be control"); 1061 Node* backedge = _head->as_Loop()->in(LoopNode::LoopBackControl); 1062 return _phase->dom_lca_internal(ctrl, backedge) == ctrl; 1063 } 1064 1065 //------------------------------add_constraint--------------------------------- 1066 // Constrain the main loop iterations so the condition: 1067 // scale_con * I + offset < limit 1068 // always holds true. That is, either increase the number of iterations in 1069 // the pre-loop or the post-loop until the condition holds true in the main 1070 // loop. Stride, scale, offset and limit are all loop invariant. Further, 1071 // stride and scale are constants (offset and limit often are). 1072 void PhaseIdealLoop::add_constraint( int stride_con, int scale_con, Node *offset, Node *limit, Node *pre_ctrl, Node **pre_limit, Node **main_limit ) { 1073 1074 // Compute "I :: (limit-offset)/scale_con" 1075 Node *con = new (C, 3) SubINode( limit, offset ); 1076 register_new_node( con, pre_ctrl ); 1077 Node *scale = _igvn.intcon(scale_con); 1078 set_ctrl(scale, C->root()); 1079 Node *X = new (C, 3) DivINode( 0, con, scale ); 1080 register_new_node( X, pre_ctrl ); 1081 1082 // For positive stride, the pre-loop limit always uses a MAX function 1083 // and the main loop a MIN function. For negative stride these are 1084 // reversed. 1085 1086 // Also for positive stride*scale the affine function is increasing, so the 1087 // pre-loop must check for underflow and the post-loop for overflow. 1088 // Negative stride*scale reverses this; pre-loop checks for overflow and 1089 // post-loop for underflow. 1090 if( stride_con*scale_con > 0 ) { 1091 // Compute I < (limit-offset)/scale_con 1092 // Adjust main-loop last iteration to be MIN/MAX(main_loop,X) 1093 *main_limit = (stride_con > 0) 1094 ? (Node*)(new (C, 3) MinINode( *main_limit, X )) 1095 : (Node*)(new (C, 3) MaxINode( *main_limit, X )); 1096 register_new_node( *main_limit, pre_ctrl ); 1097 1098 } else { 1099 // Compute (limit-offset)/scale_con + SGN(-scale_con) <= I 1100 // Add the negation of the main-loop constraint to the pre-loop. 1101 // See footnote [++] below for a derivation of the limit expression. 1102 Node *incr = _igvn.intcon(scale_con > 0 ? -1 : 1); 1103 set_ctrl(incr, C->root()); 1104 Node *adj = new (C, 3) AddINode( X, incr ); 1105 register_new_node( adj, pre_ctrl ); 1106 *pre_limit = (scale_con > 0) 1107 ? (Node*)new (C, 3) MinINode( *pre_limit, adj ) 1108 : (Node*)new (C, 3) MaxINode( *pre_limit, adj ); 1109 register_new_node( *pre_limit, pre_ctrl ); 1110 1111 // [++] Here's the algebra that justifies the pre-loop limit expression: 1112 // 1113 // NOT( scale_con * I + offset < limit ) 1114 // == 1115 // scale_con * I + offset >= limit 1116 // == 1117 // SGN(scale_con) * I >= (limit-offset)/|scale_con| 1118 // == 1119 // (limit-offset)/|scale_con| <= I * SGN(scale_con) 1120 // == 1121 // (limit-offset)/|scale_con|-1 < I * SGN(scale_con) 1122 // == 1123 // ( if (scale_con > 0) /*common case*/ 1124 // (limit-offset)/scale_con - 1 < I 1125 // else 1126 // (limit-offset)/scale_con + 1 > I 1127 // ) 1128 // ( if (scale_con > 0) /*common case*/ 1129 // (limit-offset)/scale_con + SGN(-scale_con) < I 1130 // else 1131 // (limit-offset)/scale_con + SGN(-scale_con) > I 1132 } 1133 } 1134 1135 1136 //------------------------------is_scaled_iv--------------------------------- 1137 // Return true if exp is a constant times an induction var 1138 bool PhaseIdealLoop::is_scaled_iv(Node* exp, Node* iv, int* p_scale) { 1139 if (exp == iv) { 1140 if (p_scale != NULL) { 1141 *p_scale = 1; 1142 } 1143 return true; 1144 } 1145 int opc = exp->Opcode(); 1146 if (opc == Op_MulI) { 1147 if (exp->in(1) == iv && exp->in(2)->is_Con()) { 1148 if (p_scale != NULL) { 1149 *p_scale = exp->in(2)->get_int(); 1150 } 1151 return true; 1152 } 1153 if (exp->in(2) == iv && exp->in(1)->is_Con()) { 1154 if (p_scale != NULL) { 1155 *p_scale = exp->in(1)->get_int(); 1156 } 1157 return true; 1158 } 1159 } else if (opc == Op_LShiftI) { 1160 if (exp->in(1) == iv && exp->in(2)->is_Con()) { 1161 if (p_scale != NULL) { 1162 *p_scale = 1 << exp->in(2)->get_int(); 1163 } 1164 return true; 1165 } 1166 } 1167 return false; 1168 } 1169 1170 //-----------------------------is_scaled_iv_plus_offset------------------------------ 1171 // Return true if exp is a simple induction variable expression: k1*iv + (invar + k2) 1172 bool PhaseIdealLoop::is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth) { 1173 if (is_scaled_iv(exp, iv, p_scale)) { 1174 if (p_offset != NULL) { 1175 Node *zero = _igvn.intcon(0); 1176 set_ctrl(zero, C->root()); 1177 *p_offset = zero; 1178 } 1179 return true; 1180 } 1181 int opc = exp->Opcode(); 1182 if (opc == Op_AddI) { 1183 if (is_scaled_iv(exp->in(1), iv, p_scale)) { 1184 if (p_offset != NULL) { 1185 *p_offset = exp->in(2); 1186 } 1187 return true; 1188 } 1189 if (exp->in(2)->is_Con()) { 1190 Node* offset2 = NULL; 1191 if (depth < 2 && 1192 is_scaled_iv_plus_offset(exp->in(1), iv, p_scale, 1193 p_offset != NULL ? &offset2 : NULL, depth+1)) { 1194 if (p_offset != NULL) { 1195 Node *ctrl_off2 = get_ctrl(offset2); 1196 Node* offset = new (C, 3) AddINode(offset2, exp->in(2)); 1197 register_new_node(offset, ctrl_off2); 1198 *p_offset = offset; 1199 } 1200 return true; 1201 } 1202 } 1203 } else if (opc == Op_SubI) { 1204 if (is_scaled_iv(exp->in(1), iv, p_scale)) { 1205 if (p_offset != NULL) { 1206 Node *zero = _igvn.intcon(0); 1207 set_ctrl(zero, C->root()); 1208 Node *ctrl_off = get_ctrl(exp->in(2)); 1209 Node* offset = new (C, 3) SubINode(zero, exp->in(2)); 1210 register_new_node(offset, ctrl_off); 1211 *p_offset = offset; 1212 } 1213 return true; 1214 } 1215 if (is_scaled_iv(exp->in(2), iv, p_scale)) { 1216 if (p_offset != NULL) { 1217 *p_scale *= -1; 1218 *p_offset = exp->in(1); 1219 } 1220 return true; 1221 } 1222 } 1223 return false; 1224 } 1225 1226 //------------------------------do_range_check--------------------------------- 1227 // Eliminate range-checks and other trip-counter vs loop-invariant tests. 1228 void PhaseIdealLoop::do_range_check( IdealLoopTree *loop, Node_List &old_new ) { 1229 #ifndef PRODUCT 1230 if( PrintOpto && VerifyLoopOptimizations ) { 1231 tty->print("Range Check Elimination "); 1232 loop->dump_head(); 1233 } 1234 #endif 1235 assert( RangeCheckElimination, "" ); 1236 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 1237 assert( cl->is_main_loop(), "" ); 1238 1239 // Find the trip counter; we are iteration splitting based on it 1240 Node *trip_counter = cl->phi(); 1241 // Find the main loop limit; we will trim it's iterations 1242 // to not ever trip end tests 1243 Node *main_limit = cl->limit(); 1244 // Find the pre-loop limit; we will expand it's iterations to 1245 // not ever trip low tests. 1246 Node *ctrl = cl->in(LoopNode::EntryControl); 1247 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" ); 1248 Node *iffm = ctrl->in(0); 1249 assert( iffm->Opcode() == Op_If, "" ); 1250 Node *p_f = iffm->in(0); 1251 assert( p_f->Opcode() == Op_IfFalse, "" ); 1252 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd(); 1253 assert( pre_end->loopnode()->is_pre_loop(), "" ); 1254 Node *pre_opaq1 = pre_end->limit(); 1255 // Occasionally it's possible for a pre-loop Opaque1 node to be 1256 // optimized away and then another round of loop opts attempted. 1257 // We can not optimize this particular loop in that case. 1258 if( pre_opaq1->Opcode() != Op_Opaque1 ) 1259 return; 1260 Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1; 1261 Node *pre_limit = pre_opaq->in(1); 1262 1263 // Where do we put new limit calculations 1264 Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl); 1265 1266 // Ensure the original loop limit is available from the 1267 // pre-loop Opaque1 node. 1268 Node *orig_limit = pre_opaq->original_loop_limit(); 1269 if( orig_limit == NULL || _igvn.type(orig_limit) == Type::TOP ) 1270 return; 1271 1272 // Need to find the main-loop zero-trip guard 1273 Node *bolzm = iffm->in(1); 1274 assert( bolzm->Opcode() == Op_Bool, "" ); 1275 Node *cmpzm = bolzm->in(1); 1276 assert( cmpzm->is_Cmp(), "" ); 1277 Node *opqzm = cmpzm->in(2); 1278 if( opqzm->Opcode() != Op_Opaque1 ) 1279 return; 1280 assert( opqzm->in(1) == main_limit, "do not understand situation" ); 1281 1282 // Must know if its a count-up or count-down loop 1283 1284 // protect against stride not being a constant 1285 if ( !cl->stride_is_con() ) { 1286 return; 1287 } 1288 int stride_con = cl->stride_con(); 1289 Node *zero = _igvn.intcon(0); 1290 Node *one = _igvn.intcon(1); 1291 set_ctrl(zero, C->root()); 1292 set_ctrl(one, C->root()); 1293 1294 // Range checks that do not dominate the loop backedge (ie. 1295 // conditionally executed) can lengthen the pre loop limit beyond 1296 // the original loop limit. To prevent this, the pre limit is 1297 // (for stride > 0) MINed with the original loop limit (MAXed 1298 // stride < 0) when some range_check (rc) is conditionally 1299 // executed. 1300 bool conditional_rc = false; 1301 1302 // Check loop body for tests of trip-counter plus loop-invariant vs 1303 // loop-invariant. 1304 for( uint i = 0; i < loop->_body.size(); i++ ) { 1305 Node *iff = loop->_body[i]; 1306 if( iff->Opcode() == Op_If ) { // Test? 1307 1308 // Test is an IfNode, has 2 projections. If BOTH are in the loop 1309 // we need loop unswitching instead of iteration splitting. 1310 Node *exit = loop->is_loop_exit(iff); 1311 if( !exit ) continue; 1312 int flip = (exit->Opcode() == Op_IfTrue) ? 1 : 0; 1313 1314 // Get boolean condition to test 1315 Node *i1 = iff->in(1); 1316 if( !i1->is_Bool() ) continue; 1317 BoolNode *bol = i1->as_Bool(); 1318 BoolTest b_test = bol->_test; 1319 // Flip sense of test if exit condition is flipped 1320 if( flip ) 1321 b_test = b_test.negate(); 1322 1323 // Get compare 1324 Node *cmp = bol->in(1); 1325 1326 // Look for trip_counter + offset vs limit 1327 Node *rc_exp = cmp->in(1); 1328 Node *limit = cmp->in(2); 1329 jint scale_con= 1; // Assume trip counter not scaled 1330 1331 Node *limit_c = get_ctrl(limit); 1332 if( loop->is_member(get_loop(limit_c) ) ) { 1333 // Compare might have operands swapped; commute them 1334 b_test = b_test.commute(); 1335 rc_exp = cmp->in(2); 1336 limit = cmp->in(1); 1337 limit_c = get_ctrl(limit); 1338 if( loop->is_member(get_loop(limit_c) ) ) 1339 continue; // Both inputs are loop varying; cannot RCE 1340 } 1341 // Here we know 'limit' is loop invariant 1342 1343 // 'limit' maybe pinned below the zero trip test (probably from a 1344 // previous round of rce), in which case, it can't be used in the 1345 // zero trip test expression which must occur before the zero test's if. 1346 if( limit_c == ctrl ) { 1347 continue; // Don't rce this check but continue looking for other candidates. 1348 } 1349 1350 // Check for scaled induction variable plus an offset 1351 Node *offset = NULL; 1352 1353 if (!is_scaled_iv_plus_offset(rc_exp, trip_counter, &scale_con, &offset)) { 1354 continue; 1355 } 1356 1357 Node *offset_c = get_ctrl(offset); 1358 if( loop->is_member( get_loop(offset_c) ) ) 1359 continue; // Offset is not really loop invariant 1360 // Here we know 'offset' is loop invariant. 1361 1362 // As above for the 'limit', the 'offset' maybe pinned below the 1363 // zero trip test. 1364 if( offset_c == ctrl ) { 1365 continue; // Don't rce this check but continue looking for other candidates. 1366 } 1367 1368 // At this point we have the expression as: 1369 // scale_con * trip_counter + offset :: limit 1370 // where scale_con, offset and limit are loop invariant. Trip_counter 1371 // monotonically increases by stride_con, a constant. Both (or either) 1372 // stride_con and scale_con can be negative which will flip about the 1373 // sense of the test. 1374 1375 // Adjust pre and main loop limits to guard the correct iteration set 1376 if( cmp->Opcode() == Op_CmpU ) {// Unsigned compare is really 2 tests 1377 if( b_test._test == BoolTest::lt ) { // Range checks always use lt 1378 // The overflow limit: scale*I+offset < limit 1379 add_constraint( stride_con, scale_con, offset, limit, pre_ctrl, &pre_limit, &main_limit ); 1380 // The underflow limit: 0 <= scale*I+offset. 1381 // Some math yields: -scale*I-(offset+1) < 0 1382 Node *plus_one = new (C, 3) AddINode( offset, one ); 1383 register_new_node( plus_one, pre_ctrl ); 1384 Node *neg_offset = new (C, 3) SubINode( zero, plus_one ); 1385 register_new_node( neg_offset, pre_ctrl ); 1386 add_constraint( stride_con, -scale_con, neg_offset, zero, pre_ctrl, &pre_limit, &main_limit ); 1387 if (!conditional_rc) { 1388 conditional_rc = !loop->dominates_backedge(iff); 1389 } 1390 } else { 1391 #ifndef PRODUCT 1392 if( PrintOpto ) 1393 tty->print_cr("missed RCE opportunity"); 1394 #endif 1395 continue; // In release mode, ignore it 1396 } 1397 } else { // Otherwise work on normal compares 1398 switch( b_test._test ) { 1399 case BoolTest::ge: // Convert X >= Y to -X <= -Y 1400 scale_con = -scale_con; 1401 offset = new (C, 3) SubINode( zero, offset ); 1402 register_new_node( offset, pre_ctrl ); 1403 limit = new (C, 3) SubINode( zero, limit ); 1404 register_new_node( limit, pre_ctrl ); 1405 // Fall into LE case 1406 case BoolTest::le: // Convert X <= Y to X < Y+1 1407 limit = new (C, 3) AddINode( limit, one ); 1408 register_new_node( limit, pre_ctrl ); 1409 // Fall into LT case 1410 case BoolTest::lt: 1411 add_constraint( stride_con, scale_con, offset, limit, pre_ctrl, &pre_limit, &main_limit ); 1412 if (!conditional_rc) { 1413 conditional_rc = !loop->dominates_backedge(iff); 1414 } 1415 break; 1416 default: 1417 #ifndef PRODUCT 1418 if( PrintOpto ) 1419 tty->print_cr("missed RCE opportunity"); 1420 #endif 1421 continue; // Unhandled case 1422 } 1423 } 1424 1425 // Kill the eliminated test 1426 C->set_major_progress(); 1427 Node *kill_con = _igvn.intcon( 1-flip ); 1428 set_ctrl(kill_con, C->root()); 1429 _igvn.hash_delete(iff); 1430 iff->set_req(1, kill_con); 1431 _igvn._worklist.push(iff); 1432 // Find surviving projection 1433 assert(iff->is_If(), ""); 1434 ProjNode* dp = ((IfNode*)iff)->proj_out(1-flip); 1435 // Find loads off the surviving projection; remove their control edge 1436 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) { 1437 Node* cd = dp->fast_out(i); // Control-dependent node 1438 if( cd->is_Load() ) { // Loads can now float around in the loop 1439 _igvn.hash_delete(cd); 1440 // Allow the load to float around in the loop, or before it 1441 // but NOT before the pre-loop. 1442 cd->set_req(0, ctrl); // ctrl, not NULL 1443 _igvn._worklist.push(cd); 1444 --i; 1445 --imax; 1446 } 1447 } 1448 1449 } // End of is IF 1450 1451 } 1452 1453 // Update loop limits 1454 if (conditional_rc) { 1455 pre_limit = (stride_con > 0) ? (Node*)new (C,3) MinINode(pre_limit, orig_limit) 1456 : (Node*)new (C,3) MaxINode(pre_limit, orig_limit); 1457 register_new_node(pre_limit, pre_ctrl); 1458 } 1459 _igvn.hash_delete(pre_opaq); 1460 pre_opaq->set_req(1, pre_limit); 1461 1462 // Note:: we are making the main loop limit no longer precise; 1463 // need to round up based on stride. 1464 if( stride_con != 1 && stride_con != -1 ) { // Cutout for common case 1465 // "Standard" round-up logic: ([main_limit-init+(y-1)]/y)*y+init 1466 // Hopefully, compiler will optimize for powers of 2. 1467 Node *ctrl = get_ctrl(main_limit); 1468 Node *stride = cl->stride(); 1469 Node *init = cl->init_trip(); 1470 Node *span = new (C, 3) SubINode(main_limit,init); 1471 register_new_node(span,ctrl); 1472 Node *rndup = _igvn.intcon(stride_con + ((stride_con>0)?-1:1)); 1473 Node *add = new (C, 3) AddINode(span,rndup); 1474 register_new_node(add,ctrl); 1475 Node *div = new (C, 3) DivINode(0,add,stride); 1476 register_new_node(div,ctrl); 1477 Node *mul = new (C, 3) MulINode(div,stride); 1478 register_new_node(mul,ctrl); 1479 Node *newlim = new (C, 3) AddINode(mul,init); 1480 register_new_node(newlim,ctrl); 1481 main_limit = newlim; 1482 } 1483 1484 Node *main_cle = cl->loopexit(); 1485 Node *main_bol = main_cle->in(1); 1486 // Hacking loop bounds; need private copies of exit test 1487 if( main_bol->outcnt() > 1 ) {// BoolNode shared? 1488 _igvn.hash_delete(main_cle); 1489 main_bol = main_bol->clone();// Clone a private BoolNode 1490 register_new_node( main_bol, main_cle->in(0) ); 1491 main_cle->set_req(1,main_bol); 1492 } 1493 Node *main_cmp = main_bol->in(1); 1494 if( main_cmp->outcnt() > 1 ) { // CmpNode shared? 1495 _igvn.hash_delete(main_bol); 1496 main_cmp = main_cmp->clone();// Clone a private CmpNode 1497 register_new_node( main_cmp, main_cle->in(0) ); 1498 main_bol->set_req(1,main_cmp); 1499 } 1500 // Hack the now-private loop bounds 1501 _igvn.hash_delete(main_cmp); 1502 main_cmp->set_req(2, main_limit); 1503 _igvn._worklist.push(main_cmp); 1504 // The OpaqueNode is unshared by design 1505 _igvn.hash_delete(opqzm); 1506 assert( opqzm->outcnt() == 1, "cannot hack shared node" ); 1507 opqzm->set_req(1,main_limit); 1508 _igvn._worklist.push(opqzm); 1509 } 1510 1511 //------------------------------DCE_loop_body---------------------------------- 1512 // Remove simplistic dead code from loop body 1513 void IdealLoopTree::DCE_loop_body() { 1514 for( uint i = 0; i < _body.size(); i++ ) 1515 if( _body.at(i)->outcnt() == 0 ) 1516 _body.map( i--, _body.pop() ); 1517 } 1518 1519 1520 //------------------------------adjust_loop_exit_prob-------------------------- 1521 // Look for loop-exit tests with the 50/50 (or worse) guesses from the parsing stage. 1522 // Replace with a 1-in-10 exit guess. 1523 void IdealLoopTree::adjust_loop_exit_prob( PhaseIdealLoop *phase ) { 1524 Node *test = tail(); 1525 while( test != _head ) { 1526 uint top = test->Opcode(); 1527 if( top == Op_IfTrue || top == Op_IfFalse ) { 1528 int test_con = ((ProjNode*)test)->_con; 1529 assert(top == (uint)(test_con? Op_IfTrue: Op_IfFalse), "sanity"); 1530 IfNode *iff = test->in(0)->as_If(); 1531 if( iff->outcnt() == 2 ) { // Ignore dead tests 1532 Node *bol = iff->in(1); 1533 if( bol && bol->req() > 1 && bol->in(1) && 1534 ((bol->in(1)->Opcode() == Op_StorePConditional ) || 1535 (bol->in(1)->Opcode() == Op_StoreIConditional ) || 1536 (bol->in(1)->Opcode() == Op_StoreLConditional ) || 1537 (bol->in(1)->Opcode() == Op_CompareAndSwapI ) || 1538 (bol->in(1)->Opcode() == Op_CompareAndSwapL ) || 1539 (bol->in(1)->Opcode() == Op_CompareAndSwapP ) || 1540 (bol->in(1)->Opcode() == Op_CompareAndSwapN ))) 1541 return; // Allocation loops RARELY take backedge 1542 // Find the OTHER exit path from the IF 1543 Node* ex = iff->proj_out(1-test_con); 1544 float p = iff->_prob; 1545 if( !phase->is_member( this, ex ) && iff->_fcnt == COUNT_UNKNOWN ) { 1546 if( top == Op_IfTrue ) { 1547 if( p < (PROB_FAIR + PROB_UNLIKELY_MAG(3))) { 1548 iff->_prob = PROB_STATIC_FREQUENT; 1549 } 1550 } else { 1551 if( p > (PROB_FAIR - PROB_UNLIKELY_MAG(3))) { 1552 iff->_prob = PROB_STATIC_INFREQUENT; 1553 } 1554 } 1555 } 1556 } 1557 } 1558 test = phase->idom(test); 1559 } 1560 } 1561 1562 1563 //------------------------------policy_do_remove_empty_loop-------------------- 1564 // Micro-benchmark spamming. Policy is to always remove empty loops. 1565 // The 'DO' part is to replace the trip counter with the value it will 1566 // have on the last iteration. This will break the loop. 1567 bool IdealLoopTree::policy_do_remove_empty_loop( PhaseIdealLoop *phase ) { 1568 // Minimum size must be empty loop 1569 if( _body.size() > 7/*number of nodes in an empty loop*/ ) return false; 1570 1571 if( !_head->is_CountedLoop() ) return false; // Dead loop 1572 CountedLoopNode *cl = _head->as_CountedLoop(); 1573 if( !cl->loopexit() ) return false; // Malformed loop 1574 if( !phase->is_member(this,phase->get_ctrl(cl->loopexit()->in(CountedLoopEndNode::TestValue)) ) ) 1575 return false; // Infinite loop 1576 #ifndef PRODUCT 1577 if( PrintOpto ) 1578 tty->print_cr("Removing empty loop"); 1579 #endif 1580 #ifdef ASSERT 1581 // Ensure only one phi which is the iv. 1582 Node* iv = NULL; 1583 for (DUIterator_Fast imax, i = cl->fast_outs(imax); i < imax; i++) { 1584 Node* n = cl->fast_out(i); 1585 if (n->Opcode() == Op_Phi) { 1586 assert(iv == NULL, "Too many phis" ); 1587 iv = n; 1588 } 1589 } 1590 assert(iv == cl->phi(), "Wrong phi" ); 1591 #endif 1592 // Replace the phi at loop head with the final value of the last 1593 // iteration. Then the CountedLoopEnd will collapse (backedge never 1594 // taken) and all loop-invariant uses of the exit values will be correct. 1595 Node *phi = cl->phi(); 1596 Node *final = new (phase->C, 3) SubINode( cl->limit(), cl->stride() ); 1597 phase->register_new_node(final,cl->in(LoopNode::EntryControl)); 1598 phase->_igvn.replace_node(phi,final); 1599 phase->C->set_major_progress(); 1600 return true; 1601 } 1602 1603 1604 //============================================================================= 1605 //------------------------------iteration_split_impl--------------------------- 1606 bool IdealLoopTree::iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new ) { 1607 // Check and remove empty loops (spam micro-benchmarks) 1608 if( policy_do_remove_empty_loop(phase) ) 1609 return true; // Here we removed an empty loop 1610 1611 bool should_peel = policy_peeling(phase); // Should we peel? 1612 1613 bool should_unswitch = policy_unswitching(phase); 1614 1615 // Non-counted loops may be peeled; exactly 1 iteration is peeled. 1616 // This removes loop-invariant tests (usually null checks). 1617 if( !_head->is_CountedLoop() ) { // Non-counted loop 1618 if (PartialPeelLoop && phase->partial_peel(this, old_new)) { 1619 // Partial peel succeeded so terminate this round of loop opts 1620 return false; 1621 } 1622 if( should_peel ) { // Should we peel? 1623 #ifndef PRODUCT 1624 if (PrintOpto) tty->print_cr("should_peel"); 1625 #endif 1626 phase->do_peeling(this,old_new); 1627 } else if( should_unswitch ) { 1628 phase->do_unswitching(this, old_new); 1629 } 1630 return true; 1631 } 1632 CountedLoopNode *cl = _head->as_CountedLoop(); 1633 1634 if( !cl->loopexit() ) return true; // Ignore various kinds of broken loops 1635 1636 // Do nothing special to pre- and post- loops 1637 if( cl->is_pre_loop() || cl->is_post_loop() ) return true; 1638 1639 // Compute loop trip count from profile data 1640 compute_profile_trip_cnt(phase); 1641 1642 // Before attempting fancy unrolling, RCE or alignment, see if we want 1643 // to completely unroll this loop or do loop unswitching. 1644 if( cl->is_normal_loop() ) { 1645 if (should_unswitch) { 1646 phase->do_unswitching(this, old_new); 1647 return true; 1648 } 1649 bool should_maximally_unroll = policy_maximally_unroll(phase); 1650 if( should_maximally_unroll ) { 1651 // Here we did some unrolling and peeling. Eventually we will 1652 // completely unroll this loop and it will no longer be a loop. 1653 phase->do_maximally_unroll(this,old_new); 1654 return true; 1655 } 1656 } 1657 1658 1659 // Counted loops may be peeled, may need some iterations run up 1660 // front for RCE, and may want to align loop refs to a cache 1661 // line. Thus we clone a full loop up front whose trip count is 1662 // at least 1 (if peeling), but may be several more. 1663 1664 // The main loop will start cache-line aligned with at least 1 1665 // iteration of the unrolled body (zero-trip test required) and 1666 // will have some range checks removed. 1667 1668 // A post-loop will finish any odd iterations (leftover after 1669 // unrolling), plus any needed for RCE purposes. 1670 1671 bool should_unroll = policy_unroll(phase); 1672 1673 bool should_rce = policy_range_check(phase); 1674 1675 bool should_align = policy_align(phase); 1676 1677 // If not RCE'ing (iteration splitting) or Aligning, then we do not 1678 // need a pre-loop. We may still need to peel an initial iteration but 1679 // we will not be needing an unknown number of pre-iterations. 1680 // 1681 // Basically, if may_rce_align reports FALSE first time through, 1682 // we will not be able to later do RCE or Aligning on this loop. 1683 bool may_rce_align = !policy_peel_only(phase) || should_rce || should_align; 1684 1685 // If we have any of these conditions (RCE, alignment, unrolling) met, then 1686 // we switch to the pre-/main-/post-loop model. This model also covers 1687 // peeling. 1688 if( should_rce || should_align || should_unroll ) { 1689 if( cl->is_normal_loop() ) // Convert to 'pre/main/post' loops 1690 phase->insert_pre_post_loops(this,old_new, !may_rce_align); 1691 1692 // Adjust the pre- and main-loop limits to let the pre and post loops run 1693 // with full checks, but the main-loop with no checks. Remove said 1694 // checks from the main body. 1695 if( should_rce ) 1696 phase->do_range_check(this,old_new); 1697 1698 // Double loop body for unrolling. Adjust the minimum-trip test (will do 1699 // twice as many iterations as before) and the main body limit (only do 1700 // an even number of trips). If we are peeling, we might enable some RCE 1701 // and we'd rather unroll the post-RCE'd loop SO... do not unroll if 1702 // peeling. 1703 if( should_unroll && !should_peel ) 1704 phase->do_unroll(this,old_new, true); 1705 1706 // Adjust the pre-loop limits to align the main body 1707 // iterations. 1708 if( should_align ) 1709 Unimplemented(); 1710 1711 } else { // Else we have an unchanged counted loop 1712 if( should_peel ) // Might want to peel but do nothing else 1713 phase->do_peeling(this,old_new); 1714 } 1715 return true; 1716 } 1717 1718 1719 //============================================================================= 1720 //------------------------------iteration_split-------------------------------- 1721 bool IdealLoopTree::iteration_split( PhaseIdealLoop *phase, Node_List &old_new ) { 1722 // Recursively iteration split nested loops 1723 if( _child && !_child->iteration_split( phase, old_new )) 1724 return false; 1725 1726 // Clean out prior deadwood 1727 DCE_loop_body(); 1728 1729 1730 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage. 1731 // Replace with a 1-in-10 exit guess. 1732 if( _parent /*not the root loop*/ && 1733 !_irreducible && 1734 // Also ignore the occasional dead backedge 1735 !tail()->is_top() ) { 1736 adjust_loop_exit_prob(phase); 1737 } 1738 1739 1740 // Gate unrolling, RCE and peeling efforts. 1741 if( !_child && // If not an inner loop, do not split 1742 !_irreducible && 1743 _allow_optimizations && 1744 !tail()->is_top() ) { // Also ignore the occasional dead backedge 1745 if (!_has_call) { 1746 if (!iteration_split_impl( phase, old_new )) { 1747 return false; 1748 } 1749 } else if (policy_unswitching(phase)) { 1750 phase->do_unswitching(this, old_new); 1751 } 1752 } 1753 1754 // Minor offset re-organization to remove loop-fallout uses of 1755 // trip counter. 1756 if( _head->is_CountedLoop() ) phase->reorg_offsets( this ); 1757 if( _next && !_next->iteration_split( phase, old_new )) 1758 return false; 1759 return true; 1760 } 1761 1762 //-------------------------------is_uncommon_trap_proj---------------------------- 1763 // Return true if proj is the form of "proj->[region->..]call_uct" 1764 bool PhaseIdealLoop::is_uncommon_trap_proj(ProjNode* proj, bool must_reason_predicate) { 1765 int path_limit = 10; 1766 assert(proj, "invalid argument"); 1767 Node* out = proj; 1768 for (int ct = 0; ct < path_limit; ct++) { 1769 out = out->unique_ctrl_out(); 1770 if (out == NULL || out->is_Root() || out->is_Start()) 1771 return false; 1772 if (out->is_CallStaticJava()) { 1773 int req = out->as_CallStaticJava()->uncommon_trap_request(); 1774 if (req != 0) { 1775 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(req); 1776 if (!must_reason_predicate || reason == Deoptimization::Reason_predicate){ 1777 return true; 1778 } 1779 } 1780 return false; // don't do further after call 1781 } 1782 } 1783 return false; 1784 } 1785 1786 //-------------------------------is_uncommon_trap_if_pattern------------------------- 1787 // Return true for "if(test)-> proj -> ... 1788 // | 1789 // V 1790 // other_proj->[region->..]call_uct" 1791 // 1792 // "must_reason_predicate" means the uct reason must be Reason_predicate 1793 bool PhaseIdealLoop::is_uncommon_trap_if_pattern(ProjNode *proj, bool must_reason_predicate) { 1794 Node *in0 = proj->in(0); 1795 if (!in0->is_If()) return false; 1796 // Variation of a dead If node. 1797 if (in0->outcnt() < 2) return false; 1798 IfNode* iff = in0->as_If(); 1799 1800 // we need "If(Conv2B(Opaque1(...)))" pattern for must_reason_predicate 1801 if (must_reason_predicate) { 1802 if (iff->in(1)->Opcode() != Op_Conv2B || 1803 iff->in(1)->in(1)->Opcode() != Op_Opaque1) { 1804 return false; 1805 } 1806 } 1807 1808 ProjNode* other_proj = iff->proj_out(1-proj->_con)->as_Proj(); 1809 return is_uncommon_trap_proj(other_proj, must_reason_predicate); 1810 } 1811 1812 //------------------------------create_new_if_for_predicate------------------------ 1813 // create a new if above the uct_if_pattern for the predicate to be promoted. 1814 // 1815 // before after 1816 // ---------- ---------- 1817 // ctrl ctrl 1818 // | | 1819 // | | 1820 // v v 1821 // iff new_iff 1822 // / \ / \ 1823 // / \ / \ 1824 // v v v v 1825 // uncommon_proj cont_proj if_uct if_cont 1826 // \ | | | | 1827 // \ | | | | 1828 // v v v | v 1829 // rgn loop | iff 1830 // | | / \ 1831 // | | / \ 1832 // v | v v 1833 // uncommon_trap | uncommon_proj cont_proj 1834 // \ \ | | 1835 // \ \ | | 1836 // v v v v 1837 // rgn loop 1838 // | 1839 // | 1840 // v 1841 // uncommon_trap 1842 // 1843 // 1844 // We will create a region to guard the uct call if there is no one there. 1845 // The true projecttion (if_cont) of the new_iff is returned. 1846 ProjNode* PhaseIdealLoop::create_new_if_for_predicate(ProjNode* cont_proj) { 1847 assert(is_uncommon_trap_if_pattern(cont_proj, true), "must be a uct if pattern!"); 1848 IfNode* iff = cont_proj->in(0)->as_If(); 1849 1850 ProjNode *uncommon_proj = iff->proj_out(1 - cont_proj->_con); 1851 Node *rgn = uncommon_proj->unique_ctrl_out(); 1852 assert(rgn->is_Region() || rgn->is_Call(), "must be a region or call uct"); 1853 1854 if (!rgn->is_Region()) { // create a region to guard the call 1855 assert(rgn->is_Call(), "must be call uct"); 1856 CallNode* call = rgn->as_Call(); 1857 rgn = new (C, 1) RegionNode(1); 1858 _igvn.set_type(rgn, rgn->bottom_type()); 1859 rgn->add_req(uncommon_proj); 1860 set_idom(rgn, idom(uncommon_proj), dom_depth(uncommon_proj)+1); 1861 _igvn.hash_delete(call); 1862 call->set_req(0, rgn); 1863 } 1864 1865 // Create new_iff 1866 uint iffdd = dom_depth(iff); 1867 IdealLoopTree* lp = get_loop(iff); 1868 IfNode *new_iff = new (C, 2) IfNode(iff->in(0), NULL, iff->_prob, iff->_fcnt); 1869 register_node(new_iff, lp, idom(iff), iffdd); 1870 Node *if_cont = new (C, 1) IfTrueNode(new_iff); 1871 Node *if_uct = new (C, 1) IfFalseNode(new_iff); 1872 if (cont_proj->is_IfFalse()) { 1873 // Swap 1874 Node* tmp = if_uct; if_uct = if_cont; if_cont = tmp; 1875 } 1876 register_node(if_cont, lp, new_iff, iffdd); 1877 register_node(if_uct, get_loop(rgn), new_iff, iffdd); 1878 1879 // if_cont to iff 1880 _igvn.hash_delete(iff); 1881 iff->set_req(0, if_cont); 1882 set_idom(iff, if_cont, dom_depth(iff)); 1883 1884 // if_uct to rgn 1885 _igvn.hash_delete(rgn); 1886 rgn->add_req(if_uct); 1887 Node* ridom = idom(rgn); 1888 Node* nrdom = dom_lca(ridom, new_iff); 1889 set_idom(rgn, nrdom, dom_depth(rgn)); 1890 1891 // rgn must have no phis 1892 assert(!rgn->as_Region()->has_phi(), "region must have no phis"); 1893 1894 return if_cont->as_Proj(); 1895 } 1896 1897 //------------------------------find_predicate_insertion_point-------------------------- 1898 // Find a good location to insert a predicate 1899 ProjNode* PhaseIdealLoop::find_predicate_insertion_point(Node* start_c) { 1900 if (start_c == C->root() || !start_c->is_Proj()) 1901 return NULL; 1902 if (is_uncommon_trap_if_pattern(start_c->as_Proj(), true/*Reason_Predicate*/)) { 1903 return start_c->as_Proj(); 1904 } 1905 return NULL; 1906 } 1907 1908 //------------------------------Invariance----------------------------------- 1909 // Helper class for loop_predication_impl to compute invariance on the fly and 1910 // clone invariants. 1911 class Invariance : public StackObj { 1912 VectorSet _visited, _invariant; 1913 Node_Stack _stack; 1914 VectorSet _clone_visited; 1915 Node_List _old_new; // map of old to new (clone) 1916 IdealLoopTree* _lpt; 1917 PhaseIdealLoop* _phase; 1918 1919 // Helper function to set up the invariance for invariance computation 1920 // If n is a known invariant, set up directly. Otherwise, look up the 1921 // the possibility to push n onto the stack for further processing. 1922 void visit(Node* use, Node* n) { 1923 if (_lpt->is_invariant(n)) { // known invariant 1924 _invariant.set(n->_idx); 1925 } else if (!n->is_CFG()) { 1926 Node *n_ctrl = _phase->ctrl_or_self(n); 1927 Node *u_ctrl = _phase->ctrl_or_self(use); // self if use is a CFG 1928 if (_phase->is_dominator(n_ctrl, u_ctrl)) { 1929 _stack.push(n, n->in(0) == NULL ? 1 : 0); 1930 } 1931 } 1932 } 1933 1934 // Compute invariance for "the_node" and (possibly) all its inputs recursively 1935 // on the fly 1936 void compute_invariance(Node* n) { 1937 assert(_visited.test(n->_idx), "must be"); 1938 visit(n, n); 1939 while (_stack.is_nonempty()) { 1940 Node* n = _stack.node(); 1941 uint idx = _stack.index(); 1942 if (idx == n->req()) { // all inputs are processed 1943 _stack.pop(); 1944 // n is invariant if it's inputs are all invariant 1945 bool all_inputs_invariant = true; 1946 for (uint i = 0; i < n->req(); i++) { 1947 Node* in = n->in(i); 1948 if (in == NULL) continue; 1949 assert(_visited.test(in->_idx), "must have visited input"); 1950 if (!_invariant.test(in->_idx)) { // bad guy 1951 all_inputs_invariant = false; 1952 break; 1953 } 1954 } 1955 if (all_inputs_invariant) { 1956 _invariant.set(n->_idx); // I am a invariant too 1957 } 1958 } else { // process next input 1959 _stack.set_index(idx + 1); 1960 Node* m = n->in(idx); 1961 if (m != NULL && !_visited.test_set(m->_idx)) { 1962 visit(n, m); 1963 } 1964 } 1965 } 1966 } 1967 1968 // Helper function to set up _old_new map for clone_nodes. 1969 // If n is a known invariant, set up directly ("clone" of n == n). 1970 // Otherwise, push n onto the stack for real cloning. 1971 void clone_visit(Node* n) { 1972 assert(_invariant.test(n->_idx), "must be invariant"); 1973 if (_lpt->is_invariant(n)) { // known invariant 1974 _old_new.map(n->_idx, n); 1975 } else{ // to be cloned 1976 assert (!n->is_CFG(), "should not see CFG here"); 1977 _stack.push(n, n->in(0) == NULL ? 1 : 0); 1978 } 1979 } 1980 1981 // Clone "n" and (possibly) all its inputs recursively 1982 void clone_nodes(Node* n, Node* ctrl) { 1983 clone_visit(n); 1984 while (_stack.is_nonempty()) { 1985 Node* n = _stack.node(); 1986 uint idx = _stack.index(); 1987 if (idx == n->req()) { // all inputs processed, clone n! 1988 _stack.pop(); 1989 // clone invariant node 1990 Node* n_cl = n->clone(); 1991 _old_new.map(n->_idx, n_cl); 1992 _phase->register_new_node(n_cl, ctrl); 1993 for (uint i = 0; i < n->req(); i++) { 1994 Node* in = n_cl->in(i); 1995 if (in == NULL) continue; 1996 n_cl->set_req(i, _old_new[in->_idx]); 1997 } 1998 } else { // process next input 1999 _stack.set_index(idx + 1); 2000 Node* m = n->in(idx); 2001 if (m != NULL && !_clone_visited.test_set(m->_idx)) { 2002 clone_visit(m); // visit the input 2003 } 2004 } 2005 } 2006 } 2007 2008 public: 2009 Invariance(Arena* area, IdealLoopTree* lpt) : 2010 _lpt(lpt), _phase(lpt->_phase), 2011 _visited(area), _invariant(area), _stack(area, 10 /* guess */), 2012 _clone_visited(area), _old_new(area) 2013 {} 2014 2015 // Map old to n for invariance computation and clone 2016 void map_ctrl(Node* old, Node* n) { 2017 assert(old->is_CFG() && n->is_CFG(), "must be"); 2018 _old_new.map(old->_idx, n); // "clone" of old is n 2019 _invariant.set(old->_idx); // old is invariant 2020 _clone_visited.set(old->_idx); 2021 } 2022 2023 // Driver function to compute invariance 2024 bool is_invariant(Node* n) { 2025 if (!_visited.test_set(n->_idx)) 2026 compute_invariance(n); 2027 return (_invariant.test(n->_idx) != 0); 2028 } 2029 2030 // Driver function to clone invariant 2031 Node* clone(Node* n, Node* ctrl) { 2032 assert(ctrl->is_CFG(), "must be"); 2033 assert(_invariant.test(n->_idx), "must be an invariant"); 2034 if (!_clone_visited.test(n->_idx)) 2035 clone_nodes(n, ctrl); 2036 return _old_new[n->_idx]; 2037 } 2038 }; 2039 2040 //------------------------------is_range_check_if ----------------------------------- 2041 // Returns true if the predicate of iff is in "scale*iv + offset u< load_range(ptr)" format 2042 // Note: this function is particularly designed for loop predication. We require load_range 2043 // and offset to be loop invariant computed on the fly by "invar" 2044 bool IdealLoopTree::is_range_check_if(IfNode *iff, PhaseIdealLoop *phase, Invariance& invar) const { 2045 if (!is_loop_exit(iff)) { 2046 return false; 2047 } 2048 if (!iff->in(1)->is_Bool()) { 2049 return false; 2050 } 2051 const BoolNode *bol = iff->in(1)->as_Bool(); 2052 if (bol->_test._test != BoolTest::lt) { 2053 return false; 2054 } 2055 if (!bol->in(1)->is_Cmp()) { 2056 return false; 2057 } 2058 const CmpNode *cmp = bol->in(1)->as_Cmp(); 2059 if (cmp->Opcode() != Op_CmpU ) { 2060 return false; 2061 } 2062 Node* range = cmp->in(2); 2063 if (range->Opcode() != Op_LoadRange) { 2064 const TypeInt* tint = phase->_igvn.type(range)->isa_int(); 2065 if (!OptimizeFill || tint == NULL || tint->empty() || tint->_lo < 0) { 2066 // Allow predication on positive values that aren't LoadRanges. 2067 // This allows optimization of loops where the length of the 2068 // array is a known value and doesn't need to be loaded back 2069 // from the array. 2070 return false; 2071 } 2072 } 2073 if (!invar.is_invariant(range)) { 2074 return false; 2075 } 2076 Node *iv = _head->as_CountedLoop()->phi(); 2077 int scale = 0; 2078 Node *offset = NULL; 2079 if (!phase->is_scaled_iv_plus_offset(cmp->in(1), iv, &scale, &offset)) { 2080 return false; 2081 } 2082 if(offset && !invar.is_invariant(offset)) { // offset must be invariant 2083 return false; 2084 } 2085 return true; 2086 } 2087 2088 //------------------------------rc_predicate----------------------------------- 2089 // Create a range check predicate 2090 // 2091 // for (i = init; i < limit; i += stride) { 2092 // a[scale*i+offset] 2093 // } 2094 // 2095 // Compute max(scale*i + offset) for init <= i < limit and build the predicate 2096 // as "max(scale*i + offset) u< a.length". 2097 // 2098 // There are two cases for max(scale*i + offset): 2099 // (1) stride*scale > 0 2100 // max(scale*i + offset) = scale*(limit-stride) + offset 2101 // (2) stride*scale < 0 2102 // max(scale*i + offset) = scale*init + offset 2103 BoolNode* PhaseIdealLoop::rc_predicate(Node* ctrl, 2104 int scale, Node* offset, 2105 Node* init, Node* limit, Node* stride, 2106 Node* range, bool upper) { 2107 DEBUG_ONLY(ttyLocker ttyl); 2108 if (TraceLoopPredicate) tty->print("rc_predicate "); 2109 2110 Node* max_idx_expr = init; 2111 int stride_con = stride->get_int(); 2112 if ((stride_con > 0) == (scale > 0) == upper) { 2113 max_idx_expr = new (C, 3) SubINode(limit, stride); 2114 register_new_node(max_idx_expr, ctrl); 2115 if (TraceLoopPredicate) tty->print("(limit - stride) "); 2116 } else { 2117 if (TraceLoopPredicate) tty->print("init "); 2118 } 2119 2120 if (scale != 1) { 2121 ConNode* con_scale = _igvn.intcon(scale); 2122 max_idx_expr = new (C, 3) MulINode(max_idx_expr, con_scale); 2123 register_new_node(max_idx_expr, ctrl); 2124 if (TraceLoopPredicate) tty->print("* %d ", scale); 2125 } 2126 2127 if (offset && (!offset->is_Con() || offset->get_int() != 0)){ 2128 max_idx_expr = new (C, 3) AddINode(max_idx_expr, offset); 2129 register_new_node(max_idx_expr, ctrl); 2130 if (TraceLoopPredicate) 2131 if (offset->is_Con()) tty->print("+ %d ", offset->get_int()); 2132 else tty->print("+ offset "); 2133 } 2134 2135 CmpUNode* cmp = new (C, 3) CmpUNode(max_idx_expr, range); 2136 register_new_node(cmp, ctrl); 2137 BoolNode* bol = new (C, 2) BoolNode(cmp, BoolTest::lt); 2138 register_new_node(bol, ctrl); 2139 2140 if (TraceLoopPredicate) tty->print_cr("<u range"); 2141 return bol; 2142 } 2143 2144 //------------------------------ loop_predication_impl-------------------------- 2145 // Insert loop predicates for null checks and range checks 2146 bool PhaseIdealLoop::loop_predication_impl(IdealLoopTree *loop) { 2147 if (!UseLoopPredicate) return false; 2148 2149 if (!loop->_head->is_Loop()) { 2150 // Could be a simple region when irreducible loops are present. 2151 return false; 2152 } 2153 2154 CountedLoopNode *cl = NULL; 2155 if (loop->_head->is_CountedLoop()) { 2156 cl = loop->_head->as_CountedLoop(); 2157 // do nothing for iteration-splitted loops 2158 if (!cl->is_normal_loop()) return false; 2159 } 2160 2161 // Too many traps seen? 2162 bool tmt = C->too_many_traps(C->method(), 0, Deoptimization::Reason_predicate); 2163 int tc = C->trap_count(Deoptimization::Reason_predicate); 2164 if (tmt || tc > 0) { 2165 if (TraceLoopPredicate) { 2166 tty->print_cr("too many predicate traps: %d", tc); 2167 C->method()->print(); // which method has too many predicate traps 2168 tty->print_cr(""); 2169 } 2170 return false; 2171 } 2172 2173 LoopNode *lpn = loop->_head->as_Loop(); 2174 Node* entry = lpn->in(LoopNode::EntryControl); 2175 2176 ProjNode *predicate_proj = find_predicate_insertion_point(entry); 2177 if (!predicate_proj){ 2178 #ifndef PRODUCT 2179 if (TraceLoopPredicate) { 2180 tty->print("missing predicate:"); 2181 loop->dump_head(); 2182 } 2183 #endif 2184 return false; 2185 } 2186 2187 ConNode* zero = _igvn.intcon(0); 2188 set_ctrl(zero, C->root()); 2189 Node *cond_false = new (C, 2) Conv2BNode(zero); 2190 register_new_node(cond_false, C->root()); 2191 ConNode* one = _igvn.intcon(1); 2192 set_ctrl(one, C->root()); 2193 Node *cond_true = new (C, 2) Conv2BNode(one); 2194 register_new_node(cond_true, C->root()); 2195 2196 ResourceArea *area = Thread::current()->resource_area(); 2197 Invariance invar(area, loop); 2198 2199 // Create list of if-projs such that a newer proj dominates all older 2200 // projs in the list, and they all dominate loop->tail() 2201 Node_List if_proj_list(area); 2202 LoopNode *head = loop->_head->as_Loop(); 2203 Node *current_proj = loop->tail(); //start from tail 2204 while ( current_proj != head ) { 2205 if (loop == get_loop(current_proj) && // still in the loop ? 2206 current_proj->is_Proj() && // is a projection ? 2207 current_proj->in(0)->Opcode() == Op_If) { // is a if projection ? 2208 if_proj_list.push(current_proj); 2209 } 2210 current_proj = idom(current_proj); 2211 } 2212 2213 bool hoisted = false; // true if at least one proj is promoted 2214 while (if_proj_list.size() > 0) { 2215 // Following are changed to nonnull when a predicate can be hoisted 2216 ProjNode* new_predicate_proj = NULL; 2217 2218 ProjNode* proj = if_proj_list.pop()->as_Proj(); 2219 IfNode* iff = proj->in(0)->as_If(); 2220 2221 if (!is_uncommon_trap_if_pattern(proj)) { 2222 if (loop->is_loop_exit(iff)) { 2223 // stop processing the remaining projs in the list because the execution of them 2224 // depends on the condition of "iff" (iff->in(1)). 2225 break; 2226 } else { 2227 // Both arms are inside the loop. There are two cases: 2228 // (1) there is one backward branch. In this case, any remaining proj 2229 // in the if_proj list post-dominates "iff". So, the condition of "iff" 2230 // does not determine the execution the remining projs directly, and we 2231 // can safely continue. 2232 // (2) both arms are forwarded, i.e. a diamond shape. In this case, "proj" 2233 // does not dominate loop->tail(), so it can not be in the if_proj list. 2234 continue; 2235 } 2236 } 2237 2238 Node* test = iff->in(1); 2239 if (!test->is_Bool()){ //Conv2B, ... 2240 continue; 2241 } 2242 BoolNode* bol = test->as_Bool(); 2243 if (invar.is_invariant(bol)) { 2244 // Invariant test 2245 new_predicate_proj = create_new_if_for_predicate(predicate_proj); 2246 Node* ctrl = new_predicate_proj->in(0)->as_If()->in(0); 2247 BoolNode* new_predicate_bol = invar.clone(bol, ctrl)->as_Bool(); 2248 2249 // Negate test if necessary 2250 bool negated = false; 2251 if (proj->_con != predicate_proj->_con) { 2252 new_predicate_bol = new (C, 2) BoolNode(new_predicate_bol->in(1), new_predicate_bol->_test.negate()); 2253 register_new_node(new_predicate_bol, ctrl); 2254 negated = true; 2255 } 2256 IfNode* new_predicate_iff = new_predicate_proj->in(0)->as_If(); 2257 _igvn.hash_delete(new_predicate_iff); 2258 new_predicate_iff->set_req(1, new_predicate_bol); 2259 if (TraceLoopPredicate) tty->print_cr("invariant if%s: %d", negated ? " negated" : "", new_predicate_iff->_idx); 2260 2261 } else if (cl != NULL && loop->is_range_check_if(iff, this, invar)) { 2262 assert(proj->_con == predicate_proj->_con, "must match"); 2263 2264 // Range check for counted loops 2265 const Node* cmp = bol->in(1)->as_Cmp(); 2266 Node* idx = cmp->in(1); 2267 assert(!invar.is_invariant(idx), "index is variant"); 2268 assert(cmp->in(2)->Opcode() == Op_LoadRange || OptimizeFill, "must be"); 2269 Node* rng = cmp->in(2); 2270 assert(invar.is_invariant(rng), "range must be invariant"); 2271 int scale = 1; 2272 Node* offset = zero; 2273 bool ok = is_scaled_iv_plus_offset(idx, cl->phi(), &scale, &offset); 2274 assert(ok, "must be index expression"); 2275 2276 Node* init = cl->init_trip(); 2277 Node* limit = cl->limit(); 2278 Node* stride = cl->stride(); 2279 2280 // Build if's for the upper and lower bound tests. The 2281 // lower_bound test will dominate the upper bound test and all 2282 // cloned or created nodes will use the lower bound test as 2283 // their declared control. 2284 ProjNode* lower_bound_proj = create_new_if_for_predicate(predicate_proj); 2285 ProjNode* upper_bound_proj = create_new_if_for_predicate(predicate_proj); 2286 assert(upper_bound_proj->in(0)->as_If()->in(0) == lower_bound_proj, "should dominate"); 2287 Node *ctrl = lower_bound_proj->in(0)->as_If()->in(0); 2288 2289 // Perform cloning to keep Invariance state correct since the 2290 // late schedule will place invariant things in the loop. 2291 rng = invar.clone(rng, ctrl); 2292 if (offset && offset != zero) { 2293 assert(invar.is_invariant(offset), "offset must be loop invariant"); 2294 offset = invar.clone(offset, ctrl); 2295 } 2296 2297 // Test the lower bound 2298 Node* lower_bound_bol = rc_predicate(ctrl, scale, offset, init, limit, stride, rng, false); 2299 IfNode* lower_bound_iff = lower_bound_proj->in(0)->as_If(); 2300 _igvn.hash_delete(lower_bound_iff); 2301 lower_bound_iff->set_req(1, lower_bound_bol); 2302 if (TraceLoopPredicate) tty->print_cr("lower bound check if: %d", lower_bound_iff->_idx); 2303 2304 // Test the upper bound 2305 Node* upper_bound_bol = rc_predicate(ctrl, scale, offset, init, limit, stride, rng, true); 2306 IfNode* upper_bound_iff = upper_bound_proj->in(0)->as_If(); 2307 _igvn.hash_delete(upper_bound_iff); 2308 upper_bound_iff->set_req(1, upper_bound_bol); 2309 if (TraceLoopPredicate) tty->print_cr("upper bound check if: %d", lower_bound_iff->_idx); 2310 2311 // Fall through into rest of the clean up code which will move 2312 // any dependent nodes onto the upper bound test. 2313 new_predicate_proj = upper_bound_proj; 2314 } else { 2315 // The other proj of the "iff" is a uncommon trap projection, and we can assume 2316 // the other proj will not be executed ("executed" means uct raised). 2317 continue; 2318 } 2319 2320 // Success - attach condition (new_predicate_bol) to predicate if 2321 invar.map_ctrl(proj, new_predicate_proj); // so that invariance test can be appropriate 2322 2323 // Eliminate the old if in the loop body 2324 _igvn.hash_delete(iff); 2325 iff->set_req(1, proj->is_IfFalse() ? cond_false : cond_true); 2326 2327 Node* ctrl = new_predicate_proj; // new control 2328 ProjNode* dp = proj; // old control 2329 assert(get_loop(dp) == loop, "guaranteed at the time of collecting proj"); 2330 // Find nodes (depends only on the test) off the surviving projection; 2331 // move them outside the loop with the control of proj_clone 2332 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) { 2333 Node* cd = dp->fast_out(i); // Control-dependent node 2334 if (cd->depends_only_on_test()) { 2335 assert(cd->in(0) == dp, ""); 2336 _igvn.hash_delete(cd); 2337 cd->set_req(0, ctrl); // ctrl, not NULL 2338 set_early_ctrl(cd); 2339 _igvn._worklist.push(cd); 2340 IdealLoopTree *new_loop = get_loop(get_ctrl(cd)); 2341 if (new_loop != loop) { 2342 if (!loop->_child) loop->_body.yank(cd); 2343 if (!new_loop->_child ) new_loop->_body.push(cd); 2344 } 2345 --i; 2346 --imax; 2347 } 2348 } 2349 2350 hoisted = true; 2351 C->set_major_progress(); 2352 } // end while 2353 2354 #ifndef PRODUCT 2355 // report that the loop predication has been actually performed 2356 // for this loop 2357 if (TraceLoopPredicate && hoisted) { 2358 tty->print("Loop Predication Performed:"); 2359 loop->dump_head(); 2360 } 2361 #endif 2362 2363 return hoisted; 2364 } 2365 2366 //------------------------------loop_predication-------------------------------- 2367 // driver routine for loop predication optimization 2368 bool IdealLoopTree::loop_predication( PhaseIdealLoop *phase) { 2369 bool hoisted = false; 2370 // Recursively promote predicates 2371 if ( _child ) { 2372 hoisted = _child->loop_predication( phase); 2373 } 2374 2375 // self 2376 if (!_irreducible && !tail()->is_top()) { 2377 hoisted |= phase->loop_predication_impl(this); 2378 } 2379 2380 if ( _next ) { //sibling 2381 hoisted |= _next->loop_predication( phase); 2382 } 2383 2384 return hoisted; 2385 } 2386 2387 2388 // Process all the loops in the loop tree and replace any fill 2389 // patterns with an intrisc version. 2390 bool PhaseIdealLoop::do_intrinsify_fill() { 2391 bool changed = false; 2392 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) { 2393 IdealLoopTree* lpt = iter.current(); 2394 changed |= intrinsify_fill(lpt); 2395 } 2396 return changed; 2397 } 2398 2399 2400 // Examine an inner loop looking for a a single store of an invariant 2401 // value in a unit stride loop, 2402 bool PhaseIdealLoop::match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value, 2403 Node*& shift, Node*& con) { 2404 const char* msg = NULL; 2405 Node* msg_node = NULL; 2406 2407 store_value = NULL; 2408 con = NULL; 2409 shift = NULL; 2410 2411 // Process the loop looking for stores. If there are multiple 2412 // stores or extra control flow give at this point. 2413 CountedLoopNode* head = lpt->_head->as_CountedLoop(); 2414 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) { 2415 Node* n = lpt->_body.at(i); 2416 if (n->outcnt() == 0) continue; // Ignore dead 2417 if (n->is_Store()) { 2418 if (store != NULL) { 2419 msg = "multiple stores"; 2420 break; 2421 } 2422 int opc = n->Opcode(); 2423 if (opc == Op_StoreP || opc == Op_StoreN || opc == Op_StoreCM) { 2424 msg = "oop fills not handled"; 2425 break; 2426 } 2427 Node* value = n->in(MemNode::ValueIn); 2428 if (!lpt->is_invariant(value)) { 2429 msg = "variant store value"; 2430 } else if (!_igvn.type(n->in(MemNode::Address))->isa_aryptr()) { 2431 msg = "not array address"; 2432 } 2433 store = n; 2434 store_value = value; 2435 } else if (n->is_If() && n != head->loopexit()) { 2436 msg = "extra control flow"; 2437 msg_node = n; 2438 } 2439 } 2440 2441 if (store == NULL) { 2442 // No store in loop 2443 return false; 2444 } 2445 2446 if (msg == NULL && head->stride_con() != 1) { 2447 // could handle negative strides too 2448 if (head->stride_con() < 0) { 2449 msg = "negative stride"; 2450 } else { 2451 msg = "non-unit stride"; 2452 } 2453 } 2454 2455 if (msg == NULL && !store->in(MemNode::Address)->is_AddP()) { 2456 msg = "can't handle store address"; 2457 msg_node = store->in(MemNode::Address); 2458 } 2459 2460 if (msg == NULL && 2461 (!store->in(MemNode::Memory)->is_Phi() || 2462 store->in(MemNode::Memory)->in(LoopNode::LoopBackControl) != store)) { 2463 msg = "store memory isn't proper phi"; 2464 msg_node = store->in(MemNode::Memory); 2465 } 2466 2467 // Make sure there is an appropriate fill routine 2468 BasicType t = store->as_Mem()->memory_type(); 2469 const char* fill_name; 2470 if (msg == NULL && 2471 StubRoutines::select_fill_function(t, false, fill_name) == NULL) { 2472 msg = "unsupported store"; 2473 msg_node = store; 2474 } 2475 2476 if (msg != NULL) { 2477 #ifndef PRODUCT 2478 if (TraceOptimizeFill) { 2479 tty->print_cr("not fill intrinsic candidate: %s", msg); 2480 if (msg_node != NULL) msg_node->dump(); 2481 } 2482 #endif 2483 return false; 2484 } 2485 2486 // Make sure the address expression can be handled. It should be 2487 // head->phi * elsize + con. head->phi might have a ConvI2L. 2488 Node* elements[4]; 2489 Node* conv = NULL; 2490 bool found_index = false; 2491 int count = store->in(MemNode::Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements)); 2492 for (int e = 0; e < count; e++) { 2493 Node* n = elements[e]; 2494 if (n->is_Con() && con == NULL) { 2495 con = n; 2496 } else if (n->Opcode() == Op_LShiftX && shift == NULL) { 2497 Node* value = n->in(1); 2498 #ifdef _LP64 2499 if (value->Opcode() == Op_ConvI2L) { 2500 conv = value; 2501 value = value->in(1); 2502 } 2503 #endif 2504 if (value != head->phi()) { 2505 msg = "unhandled shift in address"; 2506 } else { 2507 found_index = true; 2508 shift = n; 2509 assert(type2aelembytes(store->as_Mem()->memory_type(), true) == 1 << shift->in(2)->get_int(), "scale should match"); 2510 } 2511 } else if (n->Opcode() == Op_ConvI2L && conv == NULL) { 2512 if (n->in(1) == head->phi()) { 2513 found_index = true; 2514 conv = n; 2515 } else { 2516 msg = "unhandled input to ConvI2L"; 2517 } 2518 } else if (n == head->phi()) { 2519 // no shift, check below for allowed cases 2520 found_index = true; 2521 } else { 2522 msg = "unhandled node in address"; 2523 msg_node = n; 2524 } 2525 } 2526 2527 if (count == -1) { 2528 msg = "malformed address expression"; 2529 msg_node = store; 2530 } 2531 2532 if (!found_index) { 2533 msg = "missing use of index"; 2534 } 2535 2536 // byte sized items won't have a shift 2537 if (msg == NULL && shift == NULL && t != T_BYTE && t != T_BOOLEAN) { 2538 msg = "can't find shift"; 2539 msg_node = store; 2540 } 2541 2542 if (msg != NULL) { 2543 #ifndef PRODUCT 2544 if (TraceOptimizeFill) { 2545 tty->print_cr("not fill intrinsic: %s", msg); 2546 if (msg_node != NULL) msg_node->dump(); 2547 } 2548 #endif 2549 return false; 2550 } 2551 2552 // No make sure all the other nodes in the loop can be handled 2553 VectorSet ok(Thread::current()->resource_area()); 2554 2555 // store related values are ok 2556 ok.set(store->_idx); 2557 ok.set(store->in(MemNode::Memory)->_idx); 2558 2559 // Loop structure is ok 2560 ok.set(head->_idx); 2561 ok.set(head->loopexit()->_idx); 2562 ok.set(head->phi()->_idx); 2563 ok.set(head->incr()->_idx); 2564 ok.set(head->loopexit()->cmp_node()->_idx); 2565 ok.set(head->loopexit()->in(1)->_idx); 2566 2567 // Address elements are ok 2568 if (con) ok.set(con->_idx); 2569 if (shift) ok.set(shift->_idx); 2570 if (conv) ok.set(conv->_idx); 2571 2572 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) { 2573 Node* n = lpt->_body.at(i); 2574 if (n->outcnt() == 0) continue; // Ignore dead 2575 if (ok.test(n->_idx)) continue; 2576 // Backedge projection is ok 2577 if (n->is_IfTrue() && n->in(0) == head->loopexit()) continue; 2578 if (!n->is_AddP()) { 2579 msg = "unhandled node"; 2580 msg_node = n; 2581 break; 2582 } 2583 } 2584 2585 // Make sure no unexpected values are used outside the loop 2586 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) { 2587 Node* n = lpt->_body.at(i); 2588 // These values can be replaced with other nodes if they are used 2589 // outside the loop. 2590 if (n == store || n == head->loopexit() || n == head->incr() || n == store->in(MemNode::Memory)) continue; 2591 for (SimpleDUIterator iter(n); iter.has_next(); iter.next()) { 2592 Node* use = iter.get(); 2593 if (!lpt->_body.contains(use)) { 2594 msg = "node is used outside loop"; 2595 // lpt->_body.dump(); 2596 msg_node = n; 2597 break; 2598 } 2599 } 2600 } 2601 2602 #ifdef ASSERT 2603 if (TraceOptimizeFill) { 2604 if (msg != NULL) { 2605 tty->print_cr("no fill intrinsic: %s", msg); 2606 if (msg_node != NULL) msg_node->dump(); 2607 } else { 2608 tty->print_cr("fill intrinsic for:"); 2609 } 2610 store->dump(); 2611 if (Verbose) { 2612 lpt->_body.dump(); 2613 } 2614 } 2615 #endif 2616 2617 return msg == NULL; 2618 } 2619 2620 2621 2622 bool PhaseIdealLoop::intrinsify_fill(IdealLoopTree* lpt) { 2623 // Only for counted inner loops 2624 if (!lpt->is_counted() || !lpt->is_inner()) { 2625 return false; 2626 } 2627 2628 // Must have constant stride 2629 CountedLoopNode* head = lpt->_head->as_CountedLoop(); 2630 if (!head->stride_is_con() || !head->is_normal_loop()) { 2631 return false; 2632 } 2633 2634 // Check that the body only contains a store of a loop invariant 2635 // value that is indexed by the loop phi. 2636 Node* store = NULL; 2637 Node* store_value = NULL; 2638 Node* shift = NULL; 2639 Node* offset = NULL; 2640 if (!match_fill_loop(lpt, store, store_value, shift, offset)) { 2641 return false; 2642 } 2643 2644 // Now replace the whole loop body by a call to a fill routine that 2645 // covers the same region as the loop. 2646 Node* base = store->in(MemNode::Address)->as_AddP()->in(AddPNode::Base); 2647 2648 // Build an expression for the beginning of the copy region 2649 Node* index = head->init_trip(); 2650 #ifdef _LP64 2651 index = new (C, 2) ConvI2LNode(index); 2652 _igvn.register_new_node_with_optimizer(index); 2653 #endif 2654 if (shift != NULL) { 2655 // byte arrays don't require a shift but others do. 2656 index = new (C, 3) LShiftXNode(index, shift->in(2)); 2657 _igvn.register_new_node_with_optimizer(index); 2658 } 2659 index = new (C, 4) AddPNode(base, base, index); 2660 _igvn.register_new_node_with_optimizer(index); 2661 Node* from = new (C, 4) AddPNode(base, index, offset); 2662 _igvn.register_new_node_with_optimizer(from); 2663 // Compute the number of elements to copy 2664 Node* len = new (C, 3) SubINode(head->limit(), head->init_trip()); 2665 _igvn.register_new_node_with_optimizer(len); 2666 2667 BasicType t = store->as_Mem()->memory_type(); 2668 bool aligned = false; 2669 if (offset != NULL && head->init_trip()->is_Con()) { 2670 int element_size = type2aelembytes(t); 2671 aligned = (offset->find_intptr_t_type()->get_con() + head->init_trip()->get_int() * element_size) % HeapWordSize == 0; 2672 } 2673 2674 // Build a call to the fill routine 2675 const char* fill_name; 2676 address fill = StubRoutines::select_fill_function(t, aligned, fill_name); 2677 assert(fill != NULL, "what?"); 2678 2679 // Convert float/double to int/long for fill routines 2680 if (t == T_FLOAT) { 2681 store_value = new (C, 2) MoveF2INode(store_value); 2682 _igvn.register_new_node_with_optimizer(store_value); 2683 } else if (t == T_DOUBLE) { 2684 store_value = new (C, 2) MoveD2LNode(store_value); 2685 _igvn.register_new_node_with_optimizer(store_value); 2686 } 2687 2688 Node* mem_phi = store->in(MemNode::Memory); 2689 Node* result_ctrl; 2690 Node* result_mem; 2691 const TypeFunc* call_type = OptoRuntime::array_fill_Type(); 2692 int size = call_type->domain()->cnt(); 2693 CallLeafNode *call = new (C, size) CallLeafNoFPNode(call_type, fill, 2694 fill_name, TypeAryPtr::get_array_body_type(t)); 2695 call->init_req(TypeFunc::Parms+0, from); 2696 call->init_req(TypeFunc::Parms+1, store_value); 2697 #ifdef _LP64 2698 len = new (C, 2) ConvI2LNode(len); 2699 _igvn.register_new_node_with_optimizer(len); 2700 #endif 2701 call->init_req(TypeFunc::Parms+2, len); 2702 #ifdef _LP64 2703 call->init_req(TypeFunc::Parms+3, C->top()); 2704 #endif 2705 call->init_req( TypeFunc::Control, head->init_control()); 2706 call->init_req( TypeFunc::I_O , C->top() ) ; // does no i/o 2707 call->init_req( TypeFunc::Memory , mem_phi->in(LoopNode::EntryControl) ); 2708 call->init_req( TypeFunc::ReturnAdr, C->start()->proj_out(TypeFunc::ReturnAdr) ); 2709 call->init_req( TypeFunc::FramePtr, C->start()->proj_out(TypeFunc::FramePtr) ); 2710 _igvn.register_new_node_with_optimizer(call); 2711 result_ctrl = new (C, 1) ProjNode(call,TypeFunc::Control); 2712 _igvn.register_new_node_with_optimizer(result_ctrl); 2713 result_mem = new (C, 1) ProjNode(call,TypeFunc::Memory); 2714 _igvn.register_new_node_with_optimizer(result_mem); 2715 2716 // If this fill is tightly coupled to an allocation and overwrites 2717 // the whole body, allow it to take over the zeroing. 2718 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, this); 2719 if (alloc != NULL && alloc->is_AllocateArray()) { 2720 Node* length = alloc->as_AllocateArray()->Ideal_length(); 2721 if (head->limit() == length && 2722 head->init_trip() == _igvn.intcon(0)) { 2723 if (TraceOptimizeFill) { 2724 tty->print_cr("Eliminated zeroing in allocation"); 2725 } 2726 alloc->maybe_set_complete(&_igvn); 2727 } else { 2728 #ifdef ASSERT 2729 if (TraceOptimizeFill) { 2730 tty->print_cr("filling array but bounds don't match"); 2731 alloc->dump(); 2732 head->init_trip()->dump(); 2733 head->limit()->dump(); 2734 length->dump(); 2735 } 2736 #endif 2737 } 2738 } 2739 2740 // Redirect the old control and memory edges that are outside the loop. 2741 Node* exit = head->loopexit()->proj_out(0); 2742 // Sometimes the memory phi of the head is used as the outgoing 2743 // state of the loop. It's safe in this case to replace it with the 2744 // result_mem. 2745 _igvn.replace_node(store->in(MemNode::Memory), result_mem); 2746 _igvn.replace_node(exit, result_ctrl); 2747 _igvn.replace_node(store, result_mem); 2748 // Any uses the increment outside of the loop become the loop limit. 2749 _igvn.replace_node(head->incr(), head->limit()); 2750 2751 // Disconnect the head from the loop. 2752 for (uint i = 0; i < lpt->_body.size(); i++) { 2753 Node* n = lpt->_body.at(i); 2754 _igvn.replace_node(n, C->top()); 2755 } 2756 2757 return true; 2758 }